a16z(@a16z)

"The day SpaceX went public, historical asset allocation went out the window. Technology is the dog ...

a16z (@a16z) — Fri, 11 Sep 2026 19:30:01 GMT

"The day SpaceX went public, historical asset allocation went out the window. Technology is the dog that caught the bus."

"What was once success in venture (a unicorn) is now your average Wednesday and private companies are generating $1B in revenue on a timeline measurable in quarters."

"While we don't invest with that expectation at the outset, the majority of the portfolio can go to zero and we can still generate outperformance from a small handful of power law winners."

"It's not just a smaller number of companies generating the returns. It's a smaller number of funds with exposure to those companies."

"Winning begets winning: a fund that lands one behemoth gets the follow-on rights, the founder referrals, and the information edge that make it more likely to land the next one. Access compounds the same way outcomes do."

a16z's @jkhamehl on the case for resizing venture: a16z.news/p/catching-the…

Jen Kha@jkhamehl

x.com/i/article/2098…

🔗 View Quoted Tweet

💬3🔄1❤️11👀6978📊3

⚡ Powered by xgo.ing

Accolade Partners' Aram Verdiyan says less than 1% of venture firms have delivered consistent 3X net...

a16z (@a16z) — Fri, 11 Sep 2026 19:00:36 GMT

Accolade Partners' Aram Verdiyan says less than 1% of venture firms have delivered consistent 3X net returns over the last two decades:

"If you as an allocator have not had access to the top five to 10 companies over the last five to 10 years, you're significantly behind in terms of returns."

"We've looked at the data of 3,000 venture capital firms in the US. Only 20 have achieved consistent 3X net returns over the last two decades."

"20 companies. Less than 1%."

"What's interesting is the consistent ones consistently had access to the category-defining companies every vintage. And by the way, just having the logo is not sufficient. If you're early stage and you have a large fund, you need to own enough of it."

"Venture-like returns are possible in late stage, but your best company should be five, 10% plus of your fund. That way, you can actually return the fund on a single company. Fund-returning math in late stage didn't exist before. It now does."

"The ones that consistently have gotten access are in that top 20 out of 3,000. So if you don't have them, there is a huge dispersion of returns. And if you don't have those, you're getting the average venture return."

@aramverdi@jkhamehl@davidgeorge83

Your browser does not support the video tag.

🔗 View on Twitter

a16z@a16z

Accolade Partners' Aram Verdiyan with a16z's Jen Kha and David George on AI's extreme power law and where the next trillion dollars gets made:

The classic way to blow up a startup was throwing too much money at it. Hire a thousand people, create dueling priorities, and kill what was working.

AI turned spending into a vending machine. You put dollars in and you get something out. Money buys compute and compute alone can improve a product. Nothing has concentrated returns like this since the social networks.

The same concentration runs through the funds. Accolade looked at 3,000 US venture firms and found only 20 delivered consistent 3x net returns over two decades. What they have in common is access to the category-defining company, fund after fund.

In this conversation, Aram, Jen, and David get into why AI is being sold against labor budgets rather than software budgets, why an LP gets fired for the opposite reasons a GP does, and why getting the fund right but sizing it wrong is the same as missing it.

00 Intro
44 The compute vending machine
02 AI hit $100B in 4 years, SaaS took 15
55 Why nobody can size the TAM of AI
44 "Which layer wins" is the wrong question
44 The category winner takes it, second place takes scraps
09 3,000 venture firms, 20 with 3x returns
40 Mid-sized venture is getting squeezed out
58 Why a late stage fund needs an early stage fund
28 Why AI made picking companies harder
53 What Harvey looked like pre-reasoning
33 An LP gets fired for the opposite reason a GP does
10 Why 60 funds is too many
02 Software companies without a buyer
14 Is AI coding a head fake?
25 $12 per employee, or $7,000
32 Where bolting on AI backfires
56 The liquidity case against venture
18 The first $100 trillion company

YouTube: youtube.com/watch?v=bsdJd2…@aramverdi@AccoladePrtnrs@jkhamehl@davidgeorge83

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬8🔄0❤️11👀8883📊6

⚡ Powered by xgo.ing

Venture exposure and venture returns are hardly the same sport: - The top 1% of VC investors accoun...

a16z (@a16z) — Fri, 11 Sep 2026 18:30:02 GMT

Venture exposure and venture returns are hardly the same sport:

- The top 1% of VC investors account for 57% of the industry's net profits

- The top 5% take 90%, and the other 95% split what's left

Full piece from a16z's @jkhamehl on the power law that runs through the funds: a16z.news/p/catching-the…

Jen Kha@jkhamehl

x.com/i/article/2098…

🔗 View Quoted Tweet

💬15🔄13❤️84👀19456📊22

⚡ Powered by xgo.ing

Accolade Partners' Aram Verdiyan with a16z's Jen Kha and David George on AI's extreme power law and ...

a16z (@a16z) — Fri, 11 Sep 2026 17:56:42 GMT

Accolade Partners' Aram Verdiyan with a16z's Jen Kha and David George on AI's extreme power law and where the next trillion dollars gets made:

The classic way to blow up a startup was throwing too much money at it. Hire a thousand people, create dueling priorities, and kill what was working.

AI turned spending into a vending machine. You put dollars in and you get something out. Money buys compute and compute alone can improve a product. Nothing has concentrated returns like this since the social networks.

The same concentration runs through the funds. Accolade looked at 3,000 US venture firms and found only 20 delivered consistent 3x net returns over two decades. What they have in common is access to the category-defining company, fund after fund.

In this conversation, Aram, Jen, and David get into why AI is being sold against labor budgets rather than software budgets, why an LP gets fired for the opposite reasons a GP does, and why getting the fund right but sizing it wrong is the same as missing it.

00 Intro
44 The compute vending machine
02 AI hit $100B in 4 years, SaaS took 15
55 Why nobody can size the TAM of AI
44 "Which layer wins" is the wrong question
44 The category winner takes it, second place takes scraps
09 3,000 venture firms, 20 with 3x returns
40 Mid-sized venture is getting squeezed out
58 Why a late stage fund needs an early stage fund
28 Why AI made picking companies harder
53 What Harvey looked like pre-reasoning
33 An LP gets fired for the opposite reason a GP does
10 Why 60 funds is too many
02 Software companies without a buyer
14 Is AI coding a head fake?
25 $12 per employee, or $7,000
32 Where bolting on AI backfires
56 The liquidity case against venture
18 The first $100 trillion company

YouTube: youtube.com/watch?v=bsdJd2…@aramverdi@AccoladePrtnrs@jkhamehl@davidgeorge83

Your browser does not support the video tag.

🔗 View on Twitter

Jen Kha@jkhamehl

x.com/i/article/2098…

🔗 View Quoted Tweet

💬5🔄7❤️35👀17109📊8

⚡ Powered by xgo.ing

a16z GP Anish Acharya says three years ago, ideas were too ambitious. Today, they're too small. "In...

a16z (@a16z) — Fri, 11 Sep 2026 15:23:03 GMT

a16z GP Anish Acharya says three years ago, ideas were too ambitious. Today, they're too small.

"In the old days, which is three years ago, we would see a company, and if what they were trying to do was too ambitious, we would not engage. Just too crazy, too complex."

"Implied by that is you wouldn't do a $100 million seed because it's too much money for almost any problem. It's too much money for any person to actually manage... It doesn't make sense as an inception round."

"Today we're almost seeing the opposite problem, where an idea that's too small is not something that we want to engage with."

"You can talk about how you put $100 million to work in a seed productively. I'm not recommending you raise $100 million, but this sort of no ceiling on ambition is showing up in how we're picking companies."

"When we invest, we tell every founder, 'We're here to help you build the strongest form of your vision.' Marc told me... the way we told our story when we were raising our first fund was, 'We were going to the moon, or we were going to leave a moon-sized crater in the ground, and there was no other option.' That's how we want to work with our founders as well."

@illscience w/ @lennysan

Your browser does not support the video tag.

🔗 View on Twitter

Lenny Rachitsky@lennysan

The "permanent underclass" is a dark fantasy Silicon Valley keeps telling itself. 

@illscience (GP at @a16z) points out that by almost every measure, things have never been better, and the job-loss story hasn’t played out as people thought. Radiologists have been “cooked” for 20 years and demand keeps rising. Open engineering roles are at all-time highs. And unlike the mobile era’s winner-take-all markets, every layer of the AI stack has ~20 strong competitors.

We discuss:
🔸 Why you don’t need to be chronically online to keep up
🔸 The "/loop make me happier" opportunity in AI
🔸 Why every company is becoming a series of loops
🔸 Why moats are discovered, not designed
🔸 What's the Birkin bag version of your product?

Listen nyoutube.com/watch?v=LdIyXi…8pfsoQ

🔗 View Quoted Tweet

💬27🔄8❤️70👀21661📊25

⚡ Powered by xgo.ing

first principles all the way down

a16z (@a16z) — Thu, 10 Sep 2026 23:33:23 GMT

first principles all the way down

💬31🔄214❤️2472👀63215📊359

⚡ Powered by xgo.ing

Your employer picks your health insurance and you live with it. For the first time in decades, this...

a16z (@a16z) — Thu, 10 Sep 2026 20:36:32 GMT

Your employer picks your health insurance and you live with it.

For the first time in decades, this $1T market is opening up.

Why now? Three things are happening simultaneously:

1. Premiums keep rising, making employers want to shop for alternatives

2. DTC care raised consumer expectations, AI-native care is raising them more

3. AI is lowering the cost and complexity of building and operating a health plan

Health plans have to compete again. A market that's long felt impenetrable is now addressable for startups.

a16z’s @julesyoo on the generational replacement cycle in commercial health insurance: a16z.news/p/the-great-he…

Julie Yoo@julesyoo

x.com/i/article/2098…

🔗 View Quoted Tweet

💬19🔄14❤️173👀36744📊36

⚡ Powered by xgo.ing

We're thrilled to lead Highstock's $30M Series A. Every consumer brand makes more product than it s...

a16z (@a16z) — Thu, 10 Sep 2026 15:28:05 GMT

We're thrilled to lead Highstock's $30M Series A.

Every consumer brand makes more product than it sells – and that inventory gets stranded in warehouses, written off, or dumped into liquidation channels that cheapen the brand.

@highstock_x is an AI-powered B2B marketplace that connects brands sitting on excess inventory with a vetted network of global wholesale buyers, and handles the hard parts: matching, logistics, compliance, and payments.

Highstock now has more than $1 billion in inventory listed on the platform and works with over 100 major brands. Along the way, they’ve kept more than 10 million pounds of product out of landfills.

Congratulations, @camillevanhorne, Sean Cashin, and the Highstock team!

By @omooretweets and @venturetwins

Camille van Horne@camillevanhorne

Beyond excited to share that @highstock_x has raised a $30M Series A led by @a16z with participation from @GreylockVC, @AbstractVC, @Daybreak_Fund, and angels.

We are doubling down on our mission of creating value for brands with unproductive inventory. With this raise, we are expanding our team and building on our success in beauty to unlock a new category: apparel.

@omooretweets, @venturetwins and @jeff_jordan are the right partners to take this company to new heights - they've invested in great marketplaces like Instacart (I may be biased :) ), and have studied how AI will scale marketplace models that were previously too labor intensive to be financially viable.

Thank you to our angels: @Max, @lennysan, JJ Zhuang (Instacart) and John Adams (Instacart).

Most of all, thank you to our customers and team. We would not be here without you.

Read my reflections on our journey below.

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬11🔄7❤️112👀33864📊18

⚡ Powered by xgo.ing

We're thrilled to lead Highstock's $30M Series A. Every consumer brand makes more product than it s...

a16z (@a16z) — Thu, 10 Sep 2026 14:54:50 GMT

We're thrilled to lead Highstock's $30M Series A.

Every consumer brand makes more product than it sells – and that inventory gets stranded in warehouses, written off, or dumped into liquidation channels that cheapen the brand.

@highstock_x is an AI-powered B2B marketplace that connects brands sitting on excess inventory with a vetted network of global wholesale buyers, and handles the hard parts: matching, logistics, compliance, and payments.

Highstock now has more than $1 billion in inventory listed on the platform and works with over 100 major brands. Along the way, they’ve kept more than 10 million pounds of product out of landfills.

Congratulations, @camillevanhorne, Sean Cashin, and the Highstock team!

By @omooretweets and @venturetwins

Camille van Horne@camillevanhorne

Beyond excited to share that @highstock_x has raised a $30M Series A led by @a16z with participation from @GreylockVC, @AbstractVC, @Daybreak_Fund, and angels.

We are doubling down on our mission of creating value for brands with unproductive inventory. With this raise, we are expanding our team and building on our success in beauty to unlock a new category: apparel.

@omooretweets, @venturetwins and @jeff_jordan are the right partners to take this company to new heights - they've invested in great marketplaces like Instacart (I may be biased :) ), and have studied how AI will scale marketplace models that were previously too labor intensive to be financially viable.

Thank you to our angels: @Max, @lennysan, JJ and John.

Most of all, thank you to our customers and team. We would not be here without you.

Read my reflections on our journey below.

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬1🔄1❤️3👀3151📊1

⚡ Powered by xgo.ing

ElevenLabs co-founder and CEO Mati Staniszewski on the company's origin story, which starts with the...

a16z (@a16z) — Thu, 10 Sep 2026 14:46:11 GMT

ElevenLabs co-founder and CEO Mati Staniszewski on the company's origin story, which starts with the poorly-dubbed Polish movies he grew up watching:

"I was at Palantir, so I was helping build optimization models and bringing them to the customers. And my co-founder was at Google, and he did a lot of the text models for knowledge graphs."

"He's the smartest person I know for developing a lot of the research work, and I was happily on that intersection of building the product and bringing that to the customers."

"When we started, that was the whole goal, can we combine research and product deployment under one hood?"

"We started with audio. The actual trigger point comes from where you're from, from Poland. If you watch a movie in Polish... you have one voice narrating the whole movie. All the emotional intonation disappears."

"Everybody in Poland grows up with it. One person dubbing. And then in 2021, we experienced this still happening... The quality is poor. It's a pretty terrible experience."

"In the future, that experience will be completely different. You will have original voice, original emotions, original intonation."

"Just a week ago, we released finally a model that is able to do that extremely well and finally bring the content from one language to another."

@mati w/ @davidsenra

Your browser does not support the video tag.

🔗 View on Twitter

💬14🔄12❤️171👀34040📊34

⚡ Powered by xgo.ing

From @patrickc, examples of people quickly accomplishing ambitious things together:

a16z (@a16z) — Wed, 09 Sep 2026 22:17:34 GMT

From @patrickc, examples of people quickly accomplishing ambitious things together:

💬23🔄44❤️352👀34383📊78

⚡ Powered by xgo.ing

Vals AI CEO Rayan Krishnan on why sovereign AI makes Cold War-style verification necessary: "From m...

a16z (@a16z) — Wed, 09 Sep 2026 19:46:22 GMT

Vals AI CEO Rayan Krishnan on why sovereign AI makes Cold War-style verification necessary:

"From my very idealistic perspective, I'm surprised to see so much investment in sovereign AI."

"You could probably consolidate a lot of these efforts. But it seems like that's not the world we're in or the one we're headed towards. There are actually increased efforts to build AI in a sovereign way."

"So that takes having a shared language to communicate about what the framework for evaluations is, and where we're going to collectively align around the risks."

"There's a lot to learn from nuclear here. Reagan had this line, 'Trust but verify.' We're starting to see signs of trust in that Xi Jinping and Trump are going to be meeting... But there is no clear way to actually do the verification part of this."

"Having the shared language of evals will allow us to say things like, you have the right number of nuclear warheads. In that example, there were also flyovers, a mechanism by which a country could audit another country's nuclear stockpile."

"If there's concern about the societal or even existential risk of AI, it will necessitate us constructing this shared language of evaluations to do the verification process."

@RayanKrishnan@JenniferHli

Your browser does not support the video tag.

🔗 View on Twitter

a16z@a16z

Vals AI co-founder and CEO Rayan Krishnan with a16z's Ben Horowitz and Jennifer Li on grading AI, what it costs, and who gets to make the rules:

Every big industry eventually grows an independent testing layer. AI has credit ratings to learn from and Enron to avoid. Model capability today is still mostly self-reported.

As public benchmarks saturate and models get better at optimizing for the tests themselves, Rayan makes the case for independent, continuously evolving evaluations.

The harder problem is geopolitical. Reagan's "trust but verify" worked during the Cold War because you could fly over and count the missiles. No simple equivalent for AI models exists.

In this conversation with Erik Torenberg, they get into how you measure a model's ability to improve itself, why every good benchmark eventually has to be retired, and what happens when token spend begins to rival employee salaries.

00 Intro
20 Llama 4 on public vs private benchmarks
24 Nobody agreed how to test humans either
55 What movie ratings teach us about AI
55 The Enron problem in benchmarking
36 Why a good benchmark has to be retired
22 Evals that run for weeks, not seconds
20 Where the real workday starts at 4pm
08 A firm really is just its evals
35 Why Sonnet can cost more than Opus
42 One engineer, 6 billion tokens in a day
05 Who should set the rules for models
55 Public sector enforces, private verifies
32 Why sovereign AI is inefficient and happening anyway
00 The AI version of trust-but-verify
15 Where cyber evals have to go next

YouTube: youtube.com/watch?v=WO9c9q…@RayanKrishnan@ValsAI@bhorowitz@JenniferHli@eriktorenberg

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬10🔄5❤️66👀34638📊13

⚡ Powered by xgo.ing

Vals AI CEO Rayan Krishnan on why defensible evals are critical for enterprises' survival: "The lab...

a16z (@a16z) — Wed, 09 Sep 2026 18:25:20 GMT

Vals AI CEO Rayan Krishnan on why defensible evals are critical for enterprises' survival:

"The labs side of it is very clear. If you're raising lots of money, investing heavily in building models, it's essential for you to show why your model is getting better and why the customer should pay a premium for them."

"But what I think is still underappreciated is, on the enterprise side, this is turning out to be existential as well."

"We're in this world where it is still very unclear what ROI looks like and how to value this intelligence that's being used... Token spend may start to eclipse salary spend."

"If this is such a meaningful line item in your costs, you have to justify the ROI much more cleanly... Over time, I think a firm really is just its evals."

"The ability for a company to make its evals legible in order to solve this ROI calculus is going to be the reason why that company wins out over its competitors in the long term."

@RayanKrishnan@eriktorenberg

Your browser does not support the video tag.

🔗 View on Twitter

a16z@a16z

Vals AI co-founder and CEO Rayan Krishnan with a16z's Ben Horowitz and Jennifer Li on grading AI, what it costs, and who gets to make the rules:

Every big industry eventually grows an independent testing layer. AI has credit ratings to learn from and Enron to avoid. Model capability today is still mostly self-reported.

As public benchmarks saturate and models get better at optimizing for the tests themselves, Rayan makes the case for independent, continuously evolving evaluations.

The harder problem is geopolitical. Reagan's "trust but verify" worked during the Cold War because you could fly over and count the missiles. No simple equivalent for AI models exists.

In this conversation with Erik Torenberg, they get into how you measure a model's ability to improve itself, why every good benchmark eventually has to be retired, and what happens when token spend begins to rival employee salaries.

00 Intro
20 Llama 4 on public vs private benchmarks
24 Nobody agreed how to test humans either
55 What movie ratings teach us about AI
55 The Enron problem in benchmarking
36 Why a good benchmark has to be retired
22 Evals that run for weeks, not seconds
20 Where the real workday starts at 4pm
08 A firm really is just its evals
35 Why Sonnet can cost more than Opus
42 One engineer, 6 billion tokens in a day
05 Who should set the rules for models
55 Public sector enforces, private verifies
32 Why sovereign AI is inefficient and happening anyway
00 The AI version of trust-but-verify
15 Where cyber evals have to go next

YouTube: youtube.com/watch?v=WO9c9q…@RayanKrishnan@ValsAI@bhorowitz@JenniferHli@eriktorenberg

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬15🔄3❤️54👀33885📊14

⚡ Powered by xgo.ing

Vals AI CEO Rayan Krishnan on how they score a model's ability to build its own successor: "One ben...

a16z (@a16z) — Wed, 09 Sep 2026 17:32:38 GMT

Vals AI CEO Rayan Krishnan on how they score a model's ability to build its own successor:

"One benchmark we released recently is our Recursive Self-Improvement Index. It's a topic which a lot of the big labs have been talking about and starting to report on in their model cards."

"There isn't a shared language to talk about the RSI potential of models. So we created this as an apples-to-apples way to actually benchmark across the models."

"In an ideal world, what you want to do is take a frontier model and have it train the next version of itself and see where the delta comes from. But obviously that's very expensive and slow."

"What we're doing is forming a set of proxies for every part of the process it takes to build the next version of the model."

"There's work around pre-training, post-training, harness-level engineering, and then seeing in which mechanisms and behaviors the models are able to do very good research work and build something new, and where they're struggling."

@RayanKrishnan@JenniferHli

Your browser does not support the video tag.

🔗 View on Twitter

a16z@a16z

Vals AI co-founder and CEO Rayan Krishnan with a16z's Ben Horowitz and Jennifer Li on grading AI, what it costs, and who gets to make the rules:

Every big industry eventually grows an independent testing layer. AI has credit ratings to learn from and Enron to avoid. Model capability today is still mostly self-reported.

As public benchmarks saturate and models get better at optimizing for the tests themselves, Rayan makes the case for independent, continuously evolving evaluations.

The harder problem is geopolitical. Reagan's "trust but verify" worked during the Cold War because you could fly over and count the missiles. No simple equivalent for AI models exists.

In this conversation with Erik Torenberg, they get into how you measure a model's ability to improve itself, why every good benchmark eventually has to be retired, and what happens when token spend begins to rival employee salaries.

00 Intro
20 Llama 4 on public vs private benchmarks
24 Nobody agreed how to test humans either
55 What movie ratings teach us about AI
55 The Enron problem in benchmarking
36 Why a good benchmark has to be retired
22 Evals that run for weeks, not seconds
20 Where the real workday starts at 4pm
08 A firm really is just its evals
35 Why Sonnet can cost more than Opus
42 One engineer, 6 billion tokens in a day
05 Who should set the rules for models
55 Public sector enforces, private verifies
32 Why sovereign AI is inefficient and happening anyway
00 The AI version of trust-but-verify
15 Where cyber evals have to go next

YouTube: youtube.com/watch?v=WO9c9q…@RayanKrishnan@ValsAI@bhorowitz@JenniferHli@eriktorenberg

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬12🔄5❤️80👀34043📊14

⚡ Powered by xgo.ing

We're excited to lead Lightfield's $47M Series A. Every company runs on a CRM, and almost no one li...

a16z (@a16z) — Wed, 09 Sep 2026 16:09:39 GMT

We're excited to lead Lightfield's $47M Series A.

Every company runs on a CRM, and almost no one likes the one they have.

@lightfld is the CRM for the agentic era – a data model that needs no configuration, builds itself from your emails and calls in minutes, and fits your business rather than the other way around. Agents do the work that piles up around the record: the updates, the follow-ups, the handoffs, the reporting.

Since launching late last year, thousands of companies have adopted Lightfield.

Congrats, @keithpeiris, @hliriani, and the Lightfield team!

By @joeschmidtiv, @arampell, and @fabrisera2000

Keith Peiris@keithpeiris

Today we're announcing @lightfld's $47M Series A led by @a16z to reimagine CRM as a world model of a business.

For agents to do customer-facing work, they need to understand how your business actually works.

Salesforce wasn't designed for this. It was built 20+ years ago for humans to update records. Layer agentic processes on top and you get low quality output, because the underlying data is incomplete and lacks the structure agents need.

@lightfld updates itself from every interaction, forming a trustworthy model of your business for people and agents. Since launching last November, 5,000+ companies have signed up - and we're ripping out legacy CRM at mature organizations with hundreds of users.

Companies on Lightfield grow faster than their competitors. It finds prospects that look like your best customers and books meetings with them. It captures commitments from every conversation and automates follow-ups. It diagnoses weaknesses in your funnel and learns from your best sellers to codify what works.

The companies of the future will run on a business world model, not a Salesforce-era database. Our mission is to put that capability in the hands of every employee, and every agent, at every company.

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬28🔄19❤️273👀51706📊48

⚡ Powered by xgo.ing

Vals AI CEO Rayan Krishnan on the unlimited tokens experiment that revealed an anxious obligation fo...

a16z (@a16z) — Wed, 09 Sep 2026 16:02:58 GMT

Vals AI CEO Rayan Krishnan on the unlimited tokens experiment that revealed an anxious obligation for engineers to use models all the time, everywhere:

"I wanted to do a tokenmaxxing experiment. I was able to get unlimited access for our team for a month for some of the coding tools. We had a lot of engineers spending between one to two billion tokens a day."

"I went back and did some math, and it looked like in that month, we spent roughly $1.5 million worth of tokens... It was actually 10x more we were spending in tokens than employee salary for that month."

"We cannot continue with this mode of operation for the next month. How do we intelligently figure out what are the right tools we should use, for what teams, and what projects?"

"We found some pretty surprising insights. The Cognition Devin tool is actually very token-efficient, and so that's a place we've chosen to adopt more."

"There are a lot of places where you can get better pricing models out of subscriptions as opposed to token-based pricing."

"It's actually informed our strategy for how we can effectively tokenmaxx without spending $1.5 million per month."

@RayanKrishnan@JenniferHli

Your browser does not support the video tag.

🔗 View on Twitter

a16z@a16z

Vals AI co-founder and CEO Rayan Krishnan with a16z's Ben Horowitz and Jennifer Li on grading AI, what it costs, and who gets to make the rules:

Every big industry eventually grows an independent testing layer. AI has credit ratings to learn from and Enron to avoid. Model capability today is still mostly self-reported.

As public benchmarks saturate and models get better at optimizing for the tests themselves, Rayan makes the case for independent, continuously evolving evaluations.

The harder problem is geopolitical. Reagan's "trust but verify" worked during the Cold War because you could fly over and count the missiles. No simple equivalent for AI models exists.

In this conversation with Erik Torenberg, they get into how you measure a model's ability to improve itself, why every good benchmark eventually has to be retired, and what happens when token spend begins to rival employee salaries.

00 Intro
20 Llama 4 on public vs private benchmarks
24 Nobody agreed how to test humans either
55 What movie ratings teach us about AI
55 The Enron problem in benchmarking
36 Why a good benchmark has to be retired
22 Evals that run for weeks, not seconds
20 Where the real workday starts at 4pm
08 A firm really is just its evals
35 Why Sonnet can cost more than Opus
42 One engineer, 6 billion tokens in a day
05 Who should set the rules for models
55 Public sector enforces, private verifies
32 Why sovereign AI is inefficient and happening anyway
00 The AI version of trust-but-verify
15 Where cyber evals have to go next

YouTube: youtube.com/watch?v=WO9c9q…@RayanKrishnan@ValsAI@bhorowitz@JenniferHli@eriktorenberg

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬14🔄7❤️72👀34775📊17

⚡ Powered by xgo.ing

Vals AI co-founder and CEO Rayan Krishnan with a16z's Ben Horowitz and Jennifer Li on grading AI, wh...

a16z (@a16z) — Wed, 09 Sep 2026 14:43:26 GMT

Vals AI co-founder and CEO Rayan Krishnan with a16z's Ben Horowitz and Jennifer Li on grading AI, what it costs, and who gets to make the rules:

Every big industry eventually grows an independent testing layer. AI has credit ratings to learn from and Enron to avoid. Model capability today is still mostly self-reported.

As public benchmarks saturate and models get better at optimizing for the tests themselves, Rayan makes the case for independent, continuously evolving evaluations.

The harder problem is geopolitical. Reagan's "trust but verify" worked during the Cold War because you could fly over and count the missiles. No simple equivalent for AI models exists.

In this conversation with Erik Torenberg, they get into how you measure a model's ability to improve itself, why every good benchmark eventually has to be retired, and what happens when token spend begins to rival employee salaries.

00 Intro
20 Llama 4 on public vs private benchmarks
24 Nobody agreed how to test humans either
55 What movie ratings teach us about AI
55 The Enron problem in benchmarking
36 Why a good benchmark has to be retired
22 Evals that run for weeks, not seconds
20 Where the real workday starts at 4pm
08 A firm really is just its evals
35 Why Sonnet can cost more than Opus
42 One engineer, 6 billion tokens in a day
05 Who should set the rules for models
55 Public sector enforces, private verifies
32 Why sovereign AI is inefficient and happening anyway
00 The AI version of trust-but-verify
15 Where cyber evals have to go next

YouTube: youtube.com/watch?v=WO9c9q…@RayanKrishnan@ValsAI@bhorowitz@JenniferHli@eriktorenberg

Your browser does not support the video tag.

🔗 View on Twitter

💬22🔄14❤️140👀153524📊38

⚡ Powered by xgo.ing

We're thrilled to invest in Covenant, emerging from stealth after raising $250M. The West’s long-ra...

a16z (@a16z) — Wed, 09 Sep 2026 14:11:29 GMT

We're thrilled to invest in Covenant, emerging from stealth after raising $250M.

The West’s long-range strike capabilities cost too much to field in quantities a real conflict can demand.

@HQ_Covenant's answer is Anthem – a low-cost, long-range, heavy-payload missile designed to be manufactured at scale. Already, Anthem has flown 250+ test flights since August 2025, with government customers under contract and production lines going up in Dallas and Germany.

We’re excited to partner with @KaufmanMic, Abby Denburg, and the entire Covenant team as they make long-range defense affordable.

Michael Kaufman@KaufmanMic

Today, Covenant is emerging from stealth and introducing Anthem: a heavy-payload deep precision strike capability designed for mass production, with deliveries beginning in the thousands in 2027.

We’ve raised $250 million, flown Anthem in more than 200 tests, and secured approximately $150 million in orders, including contracts with the U.S. Army, U.S. Navy, and key international allies.

We’re building Anthem at our 105,000-square-foot production facility in Dallas, alongside factories in Germany and Israel.

Covenant is backed by @a16z, @foundersfund, @Lux_Capital, @lightspeedvp, @8vc, @aleph, and @AltimeterCap. 

Deep precision strike is the biggest gap in the arsenal of the West, and we spent the last two years building Anthem to fill it. 

Now we deliver at scale.

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬31🔄54❤️834👀165504📊102

⚡ Powered by xgo.ing

DoorDash's first office, a Stanford student housing dorm, 2013

a16z (@a16z) — Tue, 08 Sep 2026 22:22:52 GMT

DoorDash's first office, a Stanford student housing dorm, 2013

💬35🔄32❤️856👀727132📊108

⚡ Powered by xgo.ing

Mistral co-founder and CEO Arthur Mensch on open source AI and why no company will want its intellig...

a16z (@a16z) — Tue, 08 Sep 2026 19:45:25 GMT

Mistral co-founder and CEO Arthur Mensch on open source AI and why no company will want its intelligence running on someone else's grid:

"If you assume that the entire economy is going to run on AI systems, enterprises will just want to make sure that nobody can turn off their systems."

"It's the same way if you have a factory and you connect it to the grid. You want to make sure that nobody's going to turn off the grid because they don't like you... that's one of things that open source technology can bring."

"It's also the only way you can create systems that use the folklore knowledge of your employees."

"That's the knowledge you've accrued for decades. The only way to turn it into an asset that nobody else gets access to is to create your own models based on those open source models."

"This is a technology which is so important that you don't want to be locked into a single vendor."

@arthurmensch on @BigTechPod (January 2026)

Your browser does not support the video tag.

🔗 View on Twitter

💬46🔄89❤️527👀62936📊122

⚡ Powered by xgo.ing

a16z's Anish Acharya on why consumer AI is where the iPhone was in 2010, and the three reasons the b...

a16z (@a16z) — Tue, 08 Sep 2026 18:45:25 GMT

a16z's Anish Acharya on why consumer AI is where the iPhone was in 2010, and the three reasons the biggest companies in the category haven't been built yet:

"What's held us back is three things. One is that the models have been expensive, so if you want to do a free-to-use product, that hasn't been easy."

"The second is that we've kind of had an interface problem. Chat makes sense if you're the highest agency person in the world, which is Elon and Sam. But for the average consumer, their ideal interface is TikTok. So we need to find something between chat and TikTok."

"And then the technology has been so much more focused on productivity than things like human connection and entertainment."

"All those things are up for grabs. The open weight models mean things are way cheaper... Founders like @ekuyda are thinking really ambitiously about user interfaces, @bchesky from Airbnb started a foundation lab focused on next-gen user interfaces."

"I think all of those problems will get solved, or they're at least in a better position to be solved than they were two years ago."

@illscience w/ @lennysan

Your browser does not support the video tag.

🔗 View on Twitter

Lenny Rachitsky@lennysan

The "permanent underclass" is a dark fantasy Silicon Valley keeps telling itself. 

@illscience (GP at @a16z) points out that by almost every measure, things have never been better, and the job-loss story hasn’t played out as people thought. Radiologists have been “cooked” for 20 years and demand keeps rising. Open engineering roles are at all-time highs. And unlike the mobile era’s winner-take-all markets, every layer of the AI stack has ~20 strong competitors.

We discuss:
🔸 Why you don’t need to be chronically online to keep up
🔸 The "/loop make me happier" opportunity in AI
🔸 Why every company is becoming a series of loops
🔸 Why moats are discovered, not designed
🔸 What's the Birkin bag version of your product?

Listen nyoutube.com/watch?v=LdIyXi…8pfsoQ

🔗 View Quoted Tweet

💬17🔄19❤️174👀37031📊32

⚡ Powered by xgo.ing

a16z's Anish Acharya on the AI era's "permanent underclass" and the reasons he doesn't buy it: "It'...

a16z (@a16z) — Tue, 08 Sep 2026 18:00:39 GMT

a16z's Anish Acharya on the AI era's "permanent underclass" and the reasons he doesn't buy it:

"It's a funny dark fantasy that we seem to have as Silicon Valley collectively. Things have never been better, really by almost every measure."

"By how distributed all the opportunities are, by the kind of technology we have access to, to the types of ambition we're allowed to have, and to the number of companies independently working on things that are winning."

"The last era of tech was a lot more centralized... If you look at what's happening now, every part of the stack, there's not even two relevant players. There's like 20."

"You've heard all the economic data... And of course job postings are higher than they've ever been, as well as programmers. So I don't know that the empirical data bears it out."

"RSI, recursive self-improvement, is a fun term to throw around. But if you ask the most sophisticated individuals at the labs, it's not actually RSI that's occurring, which could lead to some runaway winner... It's autocatalytic effects, which just means you're using the new technology to improve your process, but it's not truly recursive."

"So everything from the most empirical to the most technical points in the other direction, and yet we can't seem to let go of this fantasy."

@illscience w/ @lennysan

Your browser does not support the video tag.

🔗 View on Twitter

Lenny Rachitsky@lennysan

The "permanent underclass" is a dark fantasy Silicon Valley keeps telling itself. 

@illscience (GP at @a16z) points out that by almost every measure, things have never been better, and the job-loss story hasn’t played out as people thought. Radiologists have been “cooked” for 20 years and demand keeps rising. Open engineering roles are at all-time highs. And unlike the mobile era’s winner-take-all markets, every layer of the AI stack has ~20 strong competitors.

We discuss:
🔸 Why you don’t need to be chronically online to keep up
🔸 The "/loop make me happier" opportunity in AI
🔸 Why every company is becoming a series of loops
🔸 Why moats are discovered, not designed
🔸 What's the Birkin bag version of your product?

Listen nyoutube.com/watch?v=LdIyXi…8pfsoQ

🔗 View Quoted Tweet

💬6🔄4❤️64👀34417📊10

⚡ Powered by xgo.ing

OpenAI's Mark Sellke and Mehtaab Sawhney on why they publish the model's mathematical thinking, not ...

a16z (@a16z) — Tue, 08 Sep 2026 16:38:26 GMT

OpenAI's Mark Sellke and Mehtaab Sawhney on why they publish the model's mathematical thinking, not just its proofs:

Mark: "We decided it was important to release these summarized chains of thought... If you've never seen these and you just see all these proofs coming out, you're not sure what it means."

"Is the model just guessing in some insane way? Is it thinking in some totally foreign way? What's going on? But actually, it's reasoning kind of shockingly like an expert human would."

Mehtaab: "It's very much like reading a colleague's notes. It's a little more disorganized in some ways."

"Especially if you work close enough with a collaborator, sometimes you'll just see them spill out their thoughts in an email to you. It feels like reading a lot of those chained together. It's quite surprising the first few times."

@MarkSellke@mehtaab_sawhney

Your browser does not support the video tag.

🔗 View on Twitter

a16z@a16z

OpenAI's Mark Sellke and Mehtaab Sawhney with a16z's Lisha Li, on the state of AI and mathematics:

Before OpenAI released GPT‑6 Astra last week, the model was already doing original mathematics. Recorded before the launch, this conversation tells the story of how it got there.

It began with GPT‑5. Mehtaab Sawhney pasted in an Erdős problem still listed as open. Five minutes later, the model surfaced a paper that had solved it. The exercise eventually uncovered published solutions to 10 more problems thought to be open.

Then Astra went further. Told to "go have fun" with a high-dimensional sphere-packing problem, it improved a bound that had stood since the 1970s. Mehtaab had spent six months on the same problem in graduate school and made "absolutely zero progress."

Another Astra result established that non-sofic groups exist with a roughly 15-page proof. A related human breakthrough took 250 pages and machinery from quantum complexity theory.

OpenAI’s Mark Sellke and Mehtaab Sawhney join a16z’s Lisha Li on why wrong ideas pollute a human’s context window, why polished papers hide how mathematics is actually made, why a breakthrough can stop one prompt early, and what math rewards once proving stops being the bottleneck.

00 Intro
44 Cracking an Erdős problem in 5 minutes
20 Why a human quits and a model doesn't
50 Wrong ideas pollute your context window
45 Why math papers are bad training data
20 Nobody knows how to stack spheres in high dimensions
28 The orange-stacking proof
04 The 1970s Russian paper nobody could beat
14 The function that won a Fields Medal
09 How Astra beat the sphere-stacking record
20 Why error correction is sphere packing in disguise
15 The breakthrough Astra almost didn't bother with
22 Solving harder problems means it has better taste
00 One model for taste, one for the grind
40 Astra found an infinite group no finite one can imitate
00 250 pages of quantum complexity, or 15 of group theory
18 Only humans write 200-page proofs
00:38 What changes when proving stops being the bottleneck
02:44 The problems AI may never solve

YouTube: youtube.com/watch?v=1JvyLG…@mehtaab_sawhney@MarkSellke@OpenAI@lishali88

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬18🔄8❤️104👀34993📊20

⚡ Powered by xgo.ing

OpenAI's Mark Sellke and Mehtaab Sawhney with a16z's Lisha Li, on the state of AI and mathematics: ...

a16z (@a16z) — Tue, 08 Sep 2026 15:16:21 GMT

OpenAI's Mark Sellke and Mehtaab Sawhney with a16z's Lisha Li, on the state of AI and mathematics:

Before OpenAI released GPT‑6 Astra last week, the model was already doing original mathematics. Recorded before the launch, this conversation tells the story of how it got there.

It began with GPT‑5. Mehtaab Sawhney pasted in an Erdős problem still listed as open. Five minutes later, the model surfaced a paper that had solved it. The exercise eventually uncovered published solutions to 10 more problems thought to be open.

Then Astra went further. Told to "go have fun" with a high-dimensional sphere-packing problem, it improved a bound that had stood since the 1970s. Mehtaab had spent six months on the same problem in graduate school and made "absolutely zero progress."

Another Astra result established that non-sofic groups exist with a roughly 15-page proof. A related human breakthrough took 250 pages and machinery from quantum complexity theory.

OpenAI’s Mark Sellke and Mehtaab Sawhney join a16z’s Lisha Li on why wrong ideas pollute a human’s context window, why polished papers hide how mathematics is actually made, why a breakthrough can stop one prompt early, and what math rewards once proving stops being the bottleneck.

00 Intro
44 Cracking an Erdős problem in 5 minutes
20 Why a human quits and a model doesn't
50 Wrong ideas pollute your context window
45 Why math papers are bad training data
20 Nobody knows how to stack spheres in high dimensions
28 The orange-stacking proof
04 The 1970s Russian paper nobody could beat
14 The function that won a Fields Medal
09 How Astra beat the sphere-stacking record
20 Why error correction is sphere packing in disguise
15 The breakthrough Astra almost didn't bother with
22 Solving harder problems means it has better taste
00 One model for taste, one for the grind
40 Astra found an infinite group no finite one can imitate
00 250 pages of quantum complexity, or 15 of group theory
18 Only humans write 200-page proofs
00:38 What changes when proving stops being the bottleneck
02:44 The problems AI may never solve

YouTube: youtube.com/watch?v=1JvyLG…@mehtaab_sawhney@MarkSellke@OpenAI@lishali88

Your browser does not support the video tag.

🔗 View on Twitter

💬16🔄19❤️128👀105381📊39

⚡ Powered by xgo.ing

Tech hiring is favoring data veterans Job postings mentioning data organization now ask for 3 more ...

a16z (@a16z) — Mon, 07 Sep 2026 17:00:08 GMT

Tech hiring is favoring data veterans

Job postings mentioning data organization now ask for 3 more years of experience than they did in 2023

6 out of the 10 biggest jumps in years of experience requirements are data related

More charts: a16z.news/p/chart-of-the…

a16z@a16z

Tech hiring is changing. Requirements per job post, since early 2025:

- Number of skills fell from ~29 to 22
- Years of experience rose from 4.6 to 4.85

Charts of the Week: a16z.news/p/chart-of-the…

🔗 View Quoted Tweet

💬19🔄28❤️228👀41815📊37

⚡ Powered by xgo.ing

The AI buildout is printing electrician and HVAC jobs Goldman attributes 300k+ construction jobs to...

a16z (@a16z) — Sun, 06 Sep 2026 21:45:01 GMT

The AI buildout is printing electrician and HVAC jobs

Goldman attributes 300k+ construction jobs to it since 2022, ~75k in the past year

These trades are growing ~2% a year, roughly double construction overall

Charts of the Week: a16z.news/p/chart-of-the…

💬53🔄140❤️832👀234485📊179

⚡ Powered by xgo.ing

The window to patch software bugs is collapsing Of the bugs hackers actually exploit, ~87% are now ...

a16z (@a16z) — Sun, 06 Sep 2026 19:30:02 GMT

The window to patch software bugs is collapsing

Of the bugs hackers actually exploit, ~87% are now being attacked on or before the day the bug is public knowledge

That share was 23% in 2020

Charts of the Week: a16z.news/p/chart-of-the…

💬69🔄113❤️817👀128261📊191

⚡ Powered by xgo.ing

World Labs co-founders Justin Johnson and Dr. Fei-Fei Li say Atlas unifies the two things computer v...

a16z (@a16z) — Sun, 06 Sep 2026 17:00:33 GMT

World Labs co-founders Justin Johnson and Dr. Fei-Fei Li say Atlas unifies the two things computer vision has always kept apart:

Justin: "Historically, reconstruction has been its own subfield in computer vision with its own specialized tasks and models. Generation is what all the text-to-video models are really good at."

"Those are great for creative applications if I want to imagine something that's never been there before."

"But now with Atlas, for the first time, we're putting these two different parts of visual intelligence together in one model. It can do both 3D reconstruction and generation together in one architecture."

"We had to make it multimodal from the start. This thing natively works on text, images, videos, and camera poses as a native input to the model... It uses 3D as a native modality that it works on."

Fei-Fei: "Computer vision has been around for more than half a century... Our field traditionally has multiple tracks. You go to a computer vision conference: you have the pixel generation track, you have some recognition track, and you have a 3D reconstruction track."

"This is an elegant model that unifies the problem of reconstruction and generation by anchoring on viewpoints and viewpoint estimation. That's just incredibly powerful."

@jcjohnss@drfeifei

Your browser does not support the video tag.

🔗 View on Twitter

a16z@a16z

World Labs co-founders Fei-Fei Li, Justin Johnson, Ben Mildenhall, and a16z's Martin Casado on Atlas, a world model for spatial intelligence:

LLMs are built on next token prediction. Video models are built on next frame prediction. Atlas is built on new view prediction, and it's the first model to unify pixel generation and pixel reconstruction, two problems computer vision has kept in separate tracks for half a century.

The practical result is a 50 to 100x reduction in what it takes to digitally capture a 3D representation of a space. Previously, you needed 100 to 300 photos of a single room. Atlas can work from just three.

In this conversation, they get into the slow motion shot from The Matrix that took hundreds of cameras and now takes three iPhones, the overnight Slack message that made them bet the company in five seconds, why robotics is bottlenecked on data rather than chips, and the case that new view prediction is AI-complete.

00 Intro
50 The Matrix slow motion scene now takes three iPhones
48 Why new view prediction is the primitive
10 Unifying generation and reconstruction
15 Gaussian splats became the bottleneck
17 Dense capture used to mean 300 photos
30 Why reconstruction needs generation to fill the gaps
44 The LLM lesson image models missed
39 The video that made them go all in
04 3D design is 95% revisions
50 The problem in robotics is data, not chips
48 Why a robot policy can't be trained like an image model
44 When the simulator becomes the planner
45 Frozen time required footage full of movement
57 Why new view prediction is AI-complete
43 Nature gave animals eyes but not trees

YouTube: youtube.com/watch?v=qn1QDD…@drfeifei@jcjohnss@BenMildenhall@theworldlabs@martin_casado

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬20🔄13❤️142👀39827📊28

⚡ Powered by xgo.ing

Cyber is having a moment Across 21 major software companies, including Apple, AWS, Microsoft, and G...

a16z (@a16z) — Sat, 05 Sep 2026 21:45:01 GMT

Cyber is having a moment

Across 21 major software companies, including Apple, AWS, Microsoft, and Google:

- Reported critical vulnerabilities never cleared 100 per month in four years
- Since spring they've jumped to over 600 per month

Charts of the Week: a16z.news/p/chart-of-the…

💬197🔄409❤️2152👀961065📊608

⚡ Powered by xgo.ing

Data center construction is going vertical Spend has jumped more than $25B in six months, roughly w...

a16z (@a16z) — Sat, 05 Sep 2026 19:30:03 GMT

Data center construction is going vertical

Spend has jumped more than $25B in six months, roughly what it gained over the previous two years combined

Construction job openings bottomed near 200k last year and are back above 300k

Charts of the Week: a16z.news/p/chart-of-the…

💬35🔄64❤️426👀128191📊87

⚡ Powered by xgo.ing

World Labs co-founders Justin Johnson and Dr. Fei-Fei Li say LLMs use next-token prediction, but spa...

a16z (@a16z) — Sat, 05 Sep 2026 17:05:35 GMT

World Labs co-founders Justin Johnson and Dr. Fei-Fei Li say LLMs use next-token prediction, but spatial intelligence has its own equivalent:

Justin: "The soft definition of AI-completeness is there's this fundamental primitive that's an AI task. But if I could solve this AI task in its full, broadest generality, it would solve any intelligence problem."

"The classic example in LLMs is that next-token prediction is AI-complete... I think from Ilya: there's a mystery novel, the thing has to read the whole novel, and the final sentence is, 'And the killer was.' Predict the next token. You could basically frame any kind of intelligence task in terms of that."

"So clearly next-token prediction is something people believe is AI-complete."

"New-view prediction, this primitive that we have in Atlas, especially generative new-view prediction, is also AI-complete."

"I want to have a world where Martin is writing a proof of the Riemann hypothesis on the blackboard, and then the camera pans over to the next whiteboard."

Fei-Fei: "Evolution had to solve new-viewpoint prediction by making animals move. Nature gave animals eyes, but nature didn't give trees eyes. Why? Because when you move, you see a new viewpoint... We do believe very strongly that next-viewpoint prediction is the equivalent of next-token prediction."

@jcjohnss@drfeifei@martin_casado

Your browser does not support the video tag.

🔗 View on Twitter

a16z@a16z

World Labs co-founders Fei-Fei Li, Justin Johnson, Ben Mildenhall, and a16z's Martin Casado on Atlas, a world model for spatial intelligence:

LLMs are built on next token prediction. Video models are built on next frame prediction. Atlas is built on new view prediction, and it's the first model to unify pixel generation and pixel reconstruction, two problems computer vision has kept in separate tracks for half a century.

The practical result is a 50 to 100x reduction in what it takes to digitally capture a 3D representation of a space. Previously, you needed 100 to 300 photos of a single room. Atlas can work from just three.

In this conversation, they get into the slow motion shot from The Matrix that took hundreds of cameras and now takes three iPhones, the overnight Slack message that made them bet the company in five seconds, why robotics is bottlenecked on data rather than chips, and the case that new view prediction is AI-complete.

00 Intro
50 The Matrix slow motion scene now takes three iPhones
48 Why new view prediction is the primitive
10 Unifying generation and reconstruction
15 Gaussian splats became the bottleneck
17 Dense capture used to mean 300 photos
30 Why reconstruction needs generation to fill the gaps
44 The LLM lesson image models missed
39 The video that made them go all in
04 3D design is 95% revisions
50 The problem in robotics is data, not chips
48 Why a robot policy can't be trained like an image model
44 When the simulator becomes the planner
45 Frozen time required footage full of movement
57 Why new view prediction is AI-complete
43 Nature gave animals eyes but not trees

YouTube: youtube.com/watch?v=qn1QDD…@drfeifei@jcjohnss@BenMildenhall@theworldlabs@martin_casado

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬24🔄31❤️354👀86877📊64

⚡ Powered by xgo.ing

Tech hiring is changing. Requirements per job post, since early 2025: - Number of skills fell from ...

a16z (@a16z) — Fri, 04 Sep 2026 21:01:01 GMT

Tech hiring is changing. Requirements per job post, since early 2025:

- Number of skills fell from ~29 to 22
- Years of experience rose from 4.6 to 4.85

Charts of the Week: a16z.news/p/chart-of-the…

💬21🔄37❤️317👀73039📊62

⚡ Powered by xgo.ing

>blogs about ray tracing for a CS project in 2012 >co-founds World Labs in 2024 >ships Atla...

a16z (@a16z) — Fri, 04 Sep 2026 20:30:04 GMT

>blogs about ray tracing for a CS project in 2012
>co-founds World Labs in 2024
>ships Atlas in 2026, a spatial intelligence model that turns a few photos into a whole 3D world

curiosity compounds

a16z@a16z

World Labs co-founders Fei-Fei Li, Justin Johnson, Ben Mildenhall, and a16z's Martin Casado on Atlas, a world model for spatial intelligence:

LLMs are built on next token prediction. Video models are built on next frame prediction. Atlas is built on new view prediction, and it's the first model to unify pixel generation and pixel reconstruction, two problems computer vision has kept in separate tracks for half a century.

The practical result is a 50 to 100x reduction in what it takes to digitally capture a 3D representation of a space. Previously, you needed 100 to 300 photos of a single room. Atlas can work from just three.

In this conversation, they get into the slow motion shot from The Matrix that took hundreds of cameras and now takes three iPhones, the overnight Slack message that made them bet the company in five seconds, why robotics is bottlenecked on data rather than chips, and the case that new view prediction is AI-complete.

00 Intro
50 The Matrix slow motion scene now takes three iPhones
48 Why new view prediction is the primitive
10 Unifying generation and reconstruction
15 Gaussian splats became the bottleneck
17 Dense capture used to mean 300 photos
30 Why reconstruction needs generation to fill the gaps
44 The LLM lesson image models missed
39 The video that made them go all in
04 3D design is 95% revisions
50 The problem in robotics is data, not chips
48 Why a robot policy can't be trained like an image model
44 When the simulator becomes the planner
45 Frozen time required footage full of movement
57 Why new view prediction is AI-complete
43 Nature gave animals eyes but not trees

YouTube: youtube.com/watch?v=qn1QDD…@drfeifei@jcjohnss@BenMildenhall@theworldlabs@martin_casado

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬13🔄22❤️353👀72017📊43

⚡ Powered by xgo.ing

>be dr. fei-fei li >born in beijing, raised in chengdu >dad moves to new jersey, follow at age 16 >s...

a16z (@a16z) — Fri, 04 Sep 2026 19:45:07 GMT

>be dr. fei-fei li
>born in beijing, raised in chengdu
>dad moves to new jersey, follow at age 16
>speak almost no english
>parents open a dry cleaner
>work the counter on weekends
>get into princeton on a full scholarship
>study physics, because why not
>caltech phd
>everyone in AI is obsessed with algorithms
>bet on data instead
>"this will never scale" they say
>do it anyway
>get 14 million images hand-labeled
>a team from toronto trains a convnet on it in 2012
>error rate falls off a cliff
>deep learning era begins
>become stanford professor
>become google cloud chief scientist
>found stanford HAI
>press starts calling you the godmother of AI
>2024, most of the field is racing on language
>you say the next frontier is spatial intelligence
>co-found World Labs in 2024
>ship Atlas in 2026, a model that turns a few photos into a whole 3D world

a16z@a16z

World Labs co-founders Fei-Fei Li, Justin Johnson, Ben Mildenhall, and a16z's Martin Casado on Atlas, a world model for spatial intelligence:

LLMs are built on next token prediction. Video models are built on next frame prediction. Atlas is built on new view prediction, and it's the first model to unify pixel generation and pixel reconstruction, two problems computer vision has kept in separate tracks for half a century.

The practical result is a 50 to 100x reduction in what it takes to digitally capture a 3D representation of a space. Previously, you needed 100 to 300 photos of a single room. Atlas can work from just three.

In this conversation, they get into the slow motion shot from The Matrix that took hundreds of cameras and now takes three iPhones, the overnight Slack message that made them bet the company in five seconds, why robotics is bottlenecked on data rather than chips, and the case that new view prediction is AI-complete.

00 Intro
50 The Matrix slow motion scene now takes three iPhones
48 Why new view prediction is the primitive
10 Unifying generation and reconstruction
15 Gaussian splats became the bottleneck
17 Dense capture used to mean 300 photos
30 Why reconstruction needs generation to fill the gaps
44 The LLM lesson image models missed
39 The video that made them go all in
04 3D design is 95% revisions
50 The problem in robotics is data, not chips
48 Why a robot policy can't be trained like an image model
44 When the simulator becomes the planner
45 Frozen time required footage full of movement
57 Why new view prediction is AI-complete
43 Nature gave animals eyes but not trees

YouTube: youtube.com/watch?v=qn1QDD…@drfeifei@jcjohnss@BenMildenhall@theworldlabs@martin_casado

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬57🔄237❤️3130👀337738📊354

⚡ Powered by xgo.ing

World Labs co-founder Justin Johnson says Atlas can recreate the famous "Bullet Time" shot from The ...

a16z (@a16z) — Fri, 04 Sep 2026 18:20:14 GMT

World Labs co-founder Justin Johnson says Atlas can recreate the famous "Bullet Time" shot from The Matrix with iPhones:

"The way they did that shot is they had a ring of hundreds of cameras. [Neo] fell over in the studio, they had hundreds of cameras viewing that angle on a green screen, and then they used those hundreds and hundreds of cameras to make that famous shot in The Matrix."

"Now with Atlas, we can do this with as few as three cameras. No studio capture, no green screen, no expensive calibration. We can literally stick three iPhones on tripods and use these to take video of something happening."

"From those three iPhone videos, we can then reframe the shot and imagine, like, freeze time, have the camera fly in as the milk is splashing up, and get these amazing frozen-time views. We can do this with just a couple cameras."

@jcjohnss

Your browser does not support the video tag.

🔗 View on Twitter

a16z@a16z

World Labs co-founders Fei-Fei Li, Justin Johnson, Ben Mildenhall, and a16z's Martin Casado on Atlas, a world model for spatial intelligence:

LLMs are built on next token prediction. Video models are built on next frame prediction. Atlas is built on new view prediction, and it's the first model to unify pixel generation and pixel reconstruction, two problems computer vision has kept in separate tracks for half a century.

The practical result is a 50 to 100x reduction in what it takes to digitally capture a 3D representation of a space. Previously, you needed 100 to 300 photos of a single room. Atlas can work from just three.

In this conversation, they get into the slow motion shot from The Matrix that took hundreds of cameras and now takes three iPhones, the overnight Slack message that made them bet the company in five seconds, why robotics is bottlenecked on data rather than chips, and the case that new view prediction is AI-complete.

00 Intro
50 The Matrix slow motion scene now takes three iPhones
48 Why new view prediction is the primitive
10 Unifying generation and reconstruction
15 Gaussian splats became the bottleneck
17 Dense capture used to mean 300 photos
30 Why reconstruction needs generation to fill the gaps
44 The LLM lesson image models missed
39 The video that made them go all in
04 3D design is 95% revisions
50 The problem in robotics is data, not chips
48 Why a robot policy can't be trained like an image model
44 When the simulator becomes the planner
45 Frozen time required footage full of movement
57 Why new view prediction is AI-complete
43 Nature gave animals eyes but not trees

YouTube: youtube.com/watch?v=qn1QDD…@drfeifei@jcjohnss@BenMildenhall@theworldlabs@martin_casado

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬14🔄22❤️176👀43106📊28

⚡ Powered by xgo.ing

Old footage sitting in your camera roll could now be reconstructable as a 3D scene. World Labs co-fo...

a16z (@a16z) — Fri, 04 Sep 2026 17:05:33 GMT

Old footage sitting in your camera roll could now be reconstructable as a 3D scene. World Labs co-founders Ben Mildenhall and Fei-Fei Li on how Atlas got there:

Ben: "In a casual sense... I took three photos of this object, or six photos of this room. I look at the photos, I can understand in my mind how those piece together. I can fill in the gaps and get it."

"But there's never really been any reconciliation between those data-driven priors and the brute force dense reconstruction, which is much more akin to scientific or medical imaging... When we say dense, we really mean dense."

"This room, I want like 100, 200, 300 photos to capture it. And what we're trying to do is bring that down to like three. We're saying like 50, 100x reduction."

"At that scale it completely flips that calculus on its head of what type of captures you reconstruct. You can go back to existing imagery you have. You can go to stuff you find on the internet and even build scenes out of that. You can go to casual videos and unearth a lot of footage that in the past we would never have treated as reconstructable, and bring it to life as 3D."

"This is something we've been playing around with a lot with Atlas. Taking old clips. I've taken a bunch of my own old captures that never worked before and put them through the system and seen a reconstruction for the first time."

Fei-Fei: "The Stanford demo is underappreciated. Anywhere between 3 to 25 images, you can reconstruct that entire Stanford quad... Everything you see is generated, but according to the laws of reconstruction. And this is really magical."

@BenMildenhall@drfeifei

Your browser does not support the video tag.

🔗 View on Twitter

a16z@a16z

World Labs co-founders Fei-Fei Li, Justin Johnson, Ben Mildenhall, and a16z's Martin Casado on Atlas, a world model for spatial intelligence:

LLMs are built on next token prediction. Video models are built on next frame prediction. Atlas is built on new view prediction, and it's the first model to unify pixel generation and pixel reconstruction, two problems computer vision has kept in separate tracks for half a century.

The practical result is a 50 to 100x reduction in what it takes to digitally capture a 3D representation of a space. Previously, you needed 100 to 300 photos of a single room. Atlas can work from just three.

In this conversation, they get into the slow motion shot from The Matrix that took hundreds of cameras and now takes three iPhones, the overnight Slack message that made them bet the company in five seconds, why robotics is bottlenecked on data rather than chips, and the case that new view prediction is AI-complete.

00 Intro
50 The Matrix slow motion scene now takes three iPhones
48 Why new view prediction is the primitive
10 Unifying generation and reconstruction
15 Gaussian splats became the bottleneck
17 Dense capture used to mean 300 photos
30 Why reconstruction needs generation to fill the gaps
44 The LLM lesson image models missed
39 The video that made them go all in
04 3D design is 95% revisions
50 The problem in robotics is data, not chips
48 Why a robot policy can't be trained like an image model
44 When the simulator becomes the planner
45 Frozen time required footage full of movement
57 Why new view prediction is AI-complete
43 Nature gave animals eyes but not trees

YouTube: youtube.com/watch?v=qn1QDD…@drfeifei@jcjohnss@BenMildenhall@theworldlabs@martin_casado

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬9🔄13❤️179👀48099📊26

⚡ Powered by xgo.ing

We're thrilled to lead a $300M investment in Gimlet Labs. AI inference is one of the fastest-growin...

a16z (@a16z) — Fri, 04 Sep 2026 16:03:45 GMT

We're thrilled to lead a $300M investment in Gimlet Labs.

AI inference is one of the fastest-growing markets in the history of capitalism, and we are running out of nearly every physical input required to serve it. Inference demand compounds at software speed; power plants, data centers, and semiconductor fabs do not.

We believe @gimletlabs has built the solution: the first multi-silicon inference cloud, designed to produce more intelligence from every watt.

This system is already delivering up to 10x gains in throughput and interactivity on frontier models within the same power envelope. In a market starved for compute, efficiency is net-new capacity and latency is product differentiation.

Wherever the existing stack ends, the Gimlet team starts. Welcome, @zainasgar, Michelle Nguyen, @oazizi, @nserrino, James Bartlett, and the Gimlet Labs team!

By @RaghuRaghuram, @sarahdingwang, @shangdaxu, @steph_zhang

Zain Asgar@zainasgar

1/ Today, we announced Gimlet Labs’ $300M Series B, led by @a16z, joined by @SapphireVC as a major investor bringing our valuation to $3B.

We started Gimlet with a simple conviction: inference would become the dominant AI workload, and its infrastructure would need to be rebuilt from the ground up.

🔗 View Quoted Tweet

💬21🔄38❤️510👀91898📊78

⚡ Powered by xgo.ing

World Labs co-founders Fei-Fei Li, Justin Johnson, Ben Mildenhall, and a16z's Martin Casado on Atlas...

a16z (@a16z) — Fri, 04 Sep 2026 14:30:54 GMT

World Labs co-founders Fei-Fei Li, Justin Johnson, Ben Mildenhall, and a16z's Martin Casado on Atlas, a world model for spatial intelligence:

LLMs are built on next token prediction. Video models are built on next frame prediction. Atlas is built on new view prediction, and it's the first model to unify pixel generation and pixel reconstruction, two problems computer vision has kept in separate tracks for half a century.

The practical result is a 50 to 100x reduction in what it takes to digitally capture a 3D representation of a space. Previously, you needed 100 to 300 photos of a single room. Atlas can work from just three.

In this conversation, they get into the slow motion shot from The Matrix that took hundreds of cameras and now takes three iPhones, the overnight Slack message that made them bet the company in five seconds, why robotics is bottlenecked on data rather than chips, and the case that new view prediction is AI-complete.

00 Intro
50 The Matrix slow motion scene now takes three iPhones
48 Why new view prediction is the primitive
10 Unifying generation and reconstruction
15 Gaussian splats became the bottleneck
17 Dense capture used to mean 300 photos
30 Why reconstruction needs generation to fill the gaps
44 The LLM lesson image models missed
39 The video that made them go all in
04 3D design is 95% revisions
50 The problem in robotics is data, not chips
48 Why a robot policy can't be trained like an image model
44 When the simulator becomes the planner
45 Frozen time required footage full of movement
57 Why new view prediction is AI-complete
43 Nature gave animals eyes but not trees

YouTube: youtube.com/watch?v=qn1QDD…@drfeifei@jcjohnss@BenMildenhall@theworldlabs@martin_casado

Your browser does not support the video tag.

🔗 View on Twitter

💬34🔄131❤️862👀730275📊222

⚡ Powered by xgo.ing

astra waiting to get added to the model selector dropdown

a16z (@a16z) — Thu, 03 Sep 2026 21:33:56 GMT

astra waiting to get added to the model selector dropdown

OpenAI@OpenAI

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬87🔄323❤️7559👀261869📊575

⚡ Powered by xgo.ing

Affirm CEO Max Levchin on the 2.5 second limit that has shaped card payments for 60 years, and how A...

a16z (@a16z) — Thu, 03 Sep 2026 20:10:17 GMT

Affirm CEO Max Levchin on the 2.5 second limit that has shaped card payments for 60 years, and how Apple and Google Pay worked around it:

"One of the things that's really amazing and subtle about Apple Pay and Google Pay and its interaction with the Visa and Mastercard networks is... Visa and Mastercard have a hard 2.5 second limit on interaction between the network, the issuing bank, and the merchant and the acquiring bank."

"And so the whole thing has to happen in 2.5 seconds, or the transaction gets retried or maybe just cancelled out, and really leaves very little room for any kind of clever innovation."

"Online you can play games. If you're doing e-commerce, you're like, well yes, we're gonna submit your card to Visa, but first we're gonna run some anti-fraud checks, and we're gonna do some other things to reduce liability. But offline, once your card is presented, 2.5 seconds, that's all you got."

"Google and Apple Pay have singularly time-shifted the whole thing by creating secure enclaves inside their chips and saying, 'I already know your card, I can do all kinds of things before I actually have to talk to Visa and Mastercard.'"

"And maybe the shocking thing is that Visa and Mastercard have not yet introduced some new standard saying, actually, it doesn't have to be 2.5 seconds at all. It could be 15 seconds while we go and get a bunch of issuers to bid on better credit quality terms for you, or any other type of innovation."

"But the sort of hard written rules of the Dee Hock era are more or less intact, and that's not 30 years, that's like more, you know, 60 years. And so I think that's probably a critical or more criticizing take on what happened."

@mlevchin

Your browser does not support the video tag.

🔗 View on Twitter

a16z@a16z

Affirm co-founders, CEO Max Levchin and a16z's Alex Rampell, on 25 years inside payments, from PayPal to agentic commerce:

Payments is the biggest market on earth. No niche in it is smaller than $100 billion, and the industry has barely moved in 60 years. Visa and Mastercard keep card transaction windows too tight to build anything new inside. Max still calls the plastic card the best payment interface ever built.

Affirm began in a 2011 email thread in which Max and Alex agreed that the nearest version of the idea was a bad business. Then a cosmetics brand using Affirm moved the offer from its checkout page to its product pages, and conversion jumped 30%.

In this conversation with Erik Torenberg, they get into why Affirm walked away from the highest margins in payments, what a negative customer acquisition cost is actually worth, and why Max is bearish on agentic shopping but bullish on agentic payments.

00 Intro
00 The fraud rule that built Apple Pay
35 Visa's problematic transaction windows
30 No niches under $100 billion
10 The email that said Affirm was a bad idea
00 The wand that should've killed the card
36 Booed off stage for presenting PayPal
02 Why crypto still can't buy a cup of coffee
16 How Max and Alex met
11 The 500-Facebook-friends idea
48 The all-nighter PHP demo
58 Every startup gets 40 years in the desert
46 Moving one line lifted conversion 30%
02 Why Affirm walked away from the highest margins in payments
03 One article created an entire industry
37 The asterisk next to 0%
17 What BNPL misses about Affirm
03 Why merchants give Affirm customers
00 Why a 3-year loan is harder, and better, than a 6-week one
19 Why PayPal produced so many founders
43 Bullish on agentic payments, bearish on agentic shopping
51 Grocery shopping is already 100% agentic

YouTube: youtube.com/watch?v=J3pegs…@mlevchin@Affirm@arampell@eriktorenberg

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬15🔄14❤️148👀38898📊30

⚡ Powered by xgo.ing

There's still whitespace for vertical AI startups. If you’re building vertical AI, look for four th...

a16z (@a16z) — Thu, 03 Sep 2026 19:20:01 GMT

There's still whitespace for vertical AI startups.

If you’re building vertical AI, look for four things in a market:

- The work repeats often enough to learn from
- Judgment matters, so there’s nuance to learn
- An expert can quickly evaluate the output and explain what to improve
- You can start with one task and grow into owning more of the job

Together, these can create a strong learning loop. That’s a key part of the vertical AI advantage. Focus can earn deeper access, more job-specific data and context, and a better understanding over time of what good work looks like.

Incumbents may own the record. Labs may own the front door. Vertical AI can still own the job.

a16z’s @seema_amble on where vertical AI can still win: a16z.news/p/the-incumben…

Seema Amble@seema_amble

In the midst of the Claudeforce announcement and the debate around systems of record, focused AI native startups can win by leaning into their data asset, depth, learning loops, and ability to work across systems and parties x.com/i/article/2095…

🔗 View Quoted Tweet

💬56🔄89❤️1070👀176603📊329

⚡ Powered by xgo.ing

Affirm CEO Max Levchin on what it was like being around Elon Musk and the rest of the PayPal Mafia i...

a16z (@a16z) — Thu, 03 Sep 2026 18:25:32 GMT

Affirm CEO Max Levchin on what it was like being around Elon Musk and the rest of the PayPal Mafia in the company's early days:

"We knew the people in the room intimately. We literally exchanged pheromones by hanging out in these sweaty rooms brainstorming, screaming at each other, spending a lot of time at whiteboards and declaring the other person's ideas idiotic."

"We knew the true version of that person. The ultra philosophical Peter Thiel, the go all in every time Elon Musk, the these days gubernatorial David Sacks."

"They all have the presentation layer, then what they're like at a dinner party, and then what they are when they're really stressed out and their head hurts and they don't know what to do."

"So as you watch one of these people go for something really big, like Elon taking humanity to Mars, which is extraordinary, you know he's a human being. I've seen him sweaty and tired and grumpy at the company kitchen a thousand times."

"They're not gods, they're humans. That inspires you to think, if that guy who's just a normal dude, I spent a lot of time with him, he's normal, and he's going for this big of an idea, I should strive too."

@mlevchin

Your browser does not support the video tag.

🔗 View on Twitter

💬16🔄81❤️1035👀159077📊175

⚡ Powered by xgo.ing

Affirm CEO Max Levchin on why the card payment interface remains unbeaten, and why better ideas fail...

a16z (@a16z) — Thu, 03 Sep 2026 17:24:34 GMT

Affirm CEO Max Levchin on why the card payment interface remains unbeaten, and why better ideas fail without reaching critical mass:

"We're forever in search of the great way to pay biometrically, and it manifests itself more in movies where people's thumbs and eyeballs get removed for authentication purposes."

"And yet we're still paying with chips. By the way, the card payment interface is the singular best user interface ever created."

"Even before PayPal there was this wand... where you could wave this wand next to a gas station and payment is secured and you can fuel your car. When I saw it I was like, this is obviously going to take over credit cards. And it didn't."

"Therein lies a really interesting lesson in payments innovation... There's not really an okay outcome in payments. You either get there and everyone needs to have your widget or your network, or it just goes into the annals of time."

"It made a lot of sense. Of course you want a little wand on your car keys, just wave and get refilled. But it's just a little bit faster, not a lot faster, than the credit card you have in your pocket that just works."

"I keep waiting for some completely different way of authenticating the payer to the device, and other than the mobile phone we haven't done better."

@mlevchin@eriktorenberg

Your browser does not support the video tag.

🔗 View on Twitter

a16z@a16z

Affirm co-founders, CEO Max Levchin and a16z's Alex Rampell, on 25 years inside payments, from PayPal to agentic commerce:

Payments is the biggest market on earth. No niche in it is smaller than $100 billion, and the industry has barely moved in 60 years. Visa and Mastercard keep card transaction windows too tight to build anything new inside. Max still calls the plastic card the best payment interface ever built.

Affirm began in a 2011 email thread in which Max and Alex agreed that the nearest version of the idea was a bad business. Then a cosmetics brand using Affirm moved the offer from its checkout page to its product pages, and conversion jumped 30%.

In this conversation with Erik Torenberg, they get into why Affirm walked away from the highest margins in payments, what a negative customer acquisition cost is actually worth, and why Max is bearish on agentic shopping but bullish on agentic payments.

00 Intro
00 The fraud rule that built Apple Pay
35 Visa's problematic transaction windows
30 No niches under $100 billion
10 The email that said Affirm was a bad idea
00 The wand that should've killed the card
36 Booed off stage for presenting PayPal
02 Why crypto still can't buy a cup of coffee
16 How Max and Alex met
11 The 500-Facebook-friends idea
48 The all-nighter PHP demo
58 Every startup gets 40 years in the desert
46 Moving one line lifted conversion 30%
02 Why Affirm walked away from the highest margins in payments
03 One article created an entire industry
37 The asterisk next to 0%
17 What BNPL misses about Affirm
03 Why merchants give Affirm customers
00 Why a 3-year loan is harder, and better, than a 6-week one
19 Why PayPal produced so many founders
43 Bullish on agentic payments, bearish on agentic shopping
51 Grocery shopping is already 100% agentic

YouTube: youtube.com/watch?v=J3pegs…@mlevchin@Affirm@arampell@eriktorenberg

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬6🔄5❤️62👀34520📊11

⚡ Powered by xgo.ing

The AI inside your business software no longer just offers to look things up for you. Salesforce upd...

a16z (@a16z) — Thu, 03 Sep 2026 16:22:26 GMT

The AI inside your business software no longer just offers to look things up for you. Salesforce updates the pipeline, Docusign redlines contacts, Workday runs HR and finance tasks.

Incumbents are upgrading from retrieval to agentic action.

a16z's Seema Amble says there's still room for startups to compete. No single record understands real work, because real work touches all systems a company runs on. Do the whole job better than anyone and you unlock a learning loop that compounds.

"The incumbent may own the record and the lab may control the front door, but the vertical AI company can still win by becoming the best at doing the job itself."

Full piece from @seema_amble on where AI-native startups still win: a16z.news/p/the-incumben…

Seema Amble@seema_amble

In the midst of the Claudeforce announcement and the debate around systems of record, focused AI native startups can win by leaning into their data asset, depth, learning loops, and ability to work across systems and parties x.com/i/article/2095…

🔗 View Quoted Tweet

💬27🔄68❤️475👀93346📊118

⚡ Powered by xgo.ing

Affirm CEO Max Levchin on why PayPal produced so many founders: "As we were interviewing certainly ...

a16z (@a16z) — Thu, 03 Sep 2026 15:26:51 GMT

Affirm CEO Max Levchin on why PayPal produced so many founders:

"As we were interviewing certainly the early team and even the later team, we kept on selecting for entrepreneurs. One of the stock questions I would ask was, 'So what are you going to do after PayPal?'"

"The answer that I welcomed the most was, 'Oh, this is the last one. I'm going to start my own afterwards.' If you're coming in here to get your final graduate degree in entrepreneurship, of course we'd love to have you."

"That's why so many people that came out of PayPal proceeded to start companies. A hot second after we all dispersed, YouTube and Yelp and... LinkedIn. That was not an accident. We were attracting entrepreneurs by design."

@mlevchin

Your browser does not support the video tag.

🔗 View on Twitter

a16z@a16z

Affirm co-founders, CEO Max Levchin and a16z's Alex Rampell, on 25 years inside payments, from PayPal to agentic commerce:

Payments is the biggest market on earth. No niche in it is smaller than $100 billion, and the industry has barely moved in 60 years. Visa and Mastercard keep card transaction windows too tight to build anything new inside. Max still calls the plastic card the best payment interface ever built.

Affirm began in a 2011 email thread in which Max and Alex agreed that the nearest version of the idea was a bad business. Then a cosmetics brand using Affirm moved the offer from its checkout page to its product pages, and conversion jumped 30%.

In this conversation with Erik Torenberg, they get into why Affirm walked away from the highest margins in payments, what a negative customer acquisition cost is actually worth, and why Max is bearish on agentic shopping but bullish on agentic payments.

00 Intro
00 The fraud rule that built Apple Pay
35 Visa's problematic transaction windows
30 No niches under $100 billion
10 The email that said Affirm was a bad idea
00 The wand that should've killed the card
36 Booed off stage for presenting PayPal
02 Why crypto still can't buy a cup of coffee
16 How Max and Alex met
11 The 500-Facebook-friends idea
48 The all-nighter PHP demo
58 Every startup gets 40 years in the desert
46 Moving one line lifted conversion 30%
02 Why Affirm walked away from the highest margins in payments
03 One article created an entire industry
37 The asterisk next to 0%
17 What BNPL misses about Affirm
03 Why merchants give Affirm customers
00 Why a 3-year loan is harder, and better, than a 6-week one
19 Why PayPal produced so many founders
43 Bullish on agentic payments, bearish on agentic shopping
51 Grocery shopping is already 100% agentic

YouTube: youtube.com/watch?v=J3pegs…@mlevchin@Affirm@arampell@eriktorenberg

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬12🔄12❤️177👀57315📊29

⚡ Powered by xgo.ing

Affirm co-founders, CEO Max Levchin and a16z's Alex Rampell, on 25 years inside payments, from PayPa...

a16z (@a16z) — Thu, 03 Sep 2026 14:34:42 GMT

Affirm co-founders, CEO Max Levchin and a16z's Alex Rampell, on 25 years inside payments, from PayPal to agentic commerce:

Payments is the biggest market on earth. No niche in it is smaller than $100 billion, and the industry has barely moved in 60 years. Visa and Mastercard keep card transaction windows too tight to build anything new inside. Max still calls the plastic card the best payment interface ever built.

Affirm began in a 2011 email thread in which Max and Alex agreed that the nearest version of the idea was a bad business. Then a cosmetics brand using Affirm moved the offer from its checkout page to its product pages, and conversion jumped 30%.

In this conversation with Erik Torenberg, they get into why Affirm walked away from the highest margins in payments, what a negative customer acquisition cost is actually worth, and why Max is bearish on agentic shopping but bullish on agentic payments.

00 Intro
00 The fraud rule that built Apple Pay
35 Visa's problematic transaction windows
30 No niches under $100 billion
10 The email that said Affirm was a bad idea
00 The wand that should've killed the card
36 Booed off stage for presenting PayPal
02 Why crypto still can't buy a cup of coffee
16 How Max and Alex met
11 The 500-Facebook-friends idea
48 The all-nighter PHP demo
58 Every startup gets 40 years in the desert
46 Moving one line lifted conversion 30%
02 Why Affirm walked away from the highest margins in payments
03 One article created an entire industry
37 The asterisk next to 0%
17 What BNPL misses about Affirm
03 Why merchants give Affirm customers
00 Why a 3-year loan is harder, and better, than a 6-week one
19 Why PayPal produced so many founders
43 Bullish on agentic payments, bearish on agentic shopping
51 Grocery shopping is already 100% agentic

YouTube: youtube.com/watch?v=J3pegs…@mlevchin@Affirm@arampell@eriktorenberg

Your browser does not support the video tag.

🔗 View on Twitter

💬24🔄31❤️308👀244537📊106

⚡ Powered by xgo.ing

World Labs co-founder Justin Johnson on how five photos from your phone could teach a robot to work ...

a16z (@a16z) — Wed, 02 Sep 2026 21:47:55 GMT

World Labs co-founder Justin Johnson on how five photos from your phone could teach a robot to work in a room it has never seen:

"I think there's a couple different tech trees that people are working on. One is this notion of fully in-context learning. Maybe I've got a robotics foundation model, then I can demonstrate to the robot once how a task should be performed, and that's enough for the robot to figure out that task."

"Another version is what we're calling real-to-sim-to-real. Maybe I've got my pre-trained robotics foundation model, but I want to adapt it to this particular environment."

"Like this studio... You could come in here and take just a couple casual videos with your phone or a couple images with your phone, use Atlas to reconstruct the space, and now stage all kinds of robotics interactions in this studio space in particular."

"Literally you could do it in five minutes. Take five photos, upload it to Atlas, have it generate a simulation, describe in natural language what kind of task you want the robot to do, have an agent go build that simulation for you, RL fine-tune your general purpose robotics foundation model, and now you've got a robot... that's perfectly adapted to this space."

@jcjohnss on @MTSlive

Your browser does not support the video tag.

🔗 View on Twitter

World Labs@theworldlabs

Introducing Atlas:

The world's first multimodal world model that generates image and video frames with pixel-perfect camera control and reconstructs them in 3D.

Model the world, move the camera, and simulate space & time.

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬14🔄23❤️156👀35832📊27

⚡ Powered by xgo.ing

Moderna CEO Stéphane Bancel on the first-ever positive Phase 3 results for an individualized mRNA ca...

a16z (@a16z) — Wed, 02 Sep 2026 17:31:23 GMT

Moderna CEO Stéphane Bancel on the first-ever positive Phase 3 results for an individualized mRNA cancer vaccine:

"After working 10 years on an individualized treatment against cancer using mRNA technology, the Phase 3 was positive. It's the first time that there is an agent in melanoma that is better than Keytruda alone."

"It's the first time there is a cancer vaccine working. The field has been doing that for 20-plus years, I think more than 1,000 clinical trials that have all failed."

"We met the primary endpoint of the study, which was recurrence-free survival... And then, to our own surprise, because this was the first interim analysis... we met the secondary endpoint, which was distant metastasis-free survival."

@JorgeCondeBio@moderna_tx

Your browser does not support the video tag.

🔗 View on Twitter

a16z@a16z

Last month Moderna reported the first individualized cancer vaccine to pass a Phase 3.

Moderna CEO Stéphane Bancel and a16z's Jorge Conde on what that actually means:

The treatment works by sequencing a patient’s tumor and healthy cells, identifying the mutations relevant to their cancer, and encoding up to 34 of them into an mRNA specific to the individual. Rather than simply unleashing the immune system, the goal is to teach it what to recognize and attack.

In this conversation, they unpack the engineering challenge of manufacturing a different medicine for every patient, how Moderna has brought the process down to roughly 42 days from biopsy to treatment, and what it would take to manufacture personalized medicines at scale.

00 Intro
35 A Phase 3 that beat the standard of care
00 Why today's immunotherapy fails 40% of patients
35 A vaccine for cancer you already have
20 Why mRNA worked where 1,000 trials failed
15 90% of cancer antigens are unique to you
35 Why this vaccine only gets better from here
00 The only thing Moderna needs is your DNA sequence
20 Why the first robot was intentionally inefficient
10 How the FDA approves a drug that's different every time
05 A cancer treatment with the side effects of a flu shot
30 Why pancreatic cancer is worth a risk
30 Rare liver disease next, then autoimmune

YouTube: youtu.be/yTNMEYeCgWw@moderna_tx@JorgeCondeBio

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬15🔄16❤️122👀34651📊23

⚡ Powered by xgo.ing

Last month Moderna reported the first individualized cancer vaccine to pass a Phase 3. Moderna CEO ...

a16z (@a16z) — Wed, 02 Sep 2026 16:00:01 GMT

Last month Moderna reported the first individualized cancer vaccine to pass a Phase 3.

Moderna CEO Stéphane Bancel and a16z's Jorge Conde on what that actually means:

The treatment works by sequencing a patient’s tumor and healthy cells, identifying the mutations relevant to their cancer, and encoding up to 34 of them into an mRNA specific to the individual. Rather than simply unleashing the immune system, the goal is to teach it what to recognize and attack.

In this conversation, they unpack the engineering challenge of manufacturing a different medicine for every patient, how Moderna has brought the process down to roughly 42 days from biopsy to treatment, and what it would take to manufacture personalized medicines at scale.

00 Intro
35 A Phase 3 that beat the standard of care
00 Why today's immunotherapy fails 40% of patients
35 A vaccine for cancer you already have
20 Why mRNA worked where 1,000 trials failed
15 90% of cancer antigens are unique to you
35 Why this vaccine only gets better from here
00 The only thing Moderna needs is your DNA sequence
20 Why the first robot was intentionally inefficient
10 How the FDA approves a drug that's different every time
05 A cancer treatment with the side effects of a flu shot
30 Why pancreatic cancer is worth a risk
30 Rare liver disease next, then autoimmune

YouTube: youtu.be/yTNMEYeCgWw@moderna_tx@JorgeCondeBio

Your browser does not support the video tag.

🔗 View on Twitter

💬16🔄21❤️129👀69223📊37

⚡ Powered by xgo.ing

a16z (@a16z) — Wed, 02 Sep 2026 00:45:03 GMT

💬239🔄2751❤️25615👀1793217📊3216