elvis(@omarsar0)

Build and own your harness, folks. Very few people understand the magic behind customizing and opt...

elvis (@omarsar0) — Fri, 11 Sep 2026 16:58:55 GMT

Build and own your harness, folks. 

Very few people understand the magic behind customizing and optimizing an agent harness for your work. 

You can start seeing incredible results quickly, even with a minimal harness. Just better code, better outputs, better costs, better writing, and so on. 

As an example, I see way too many people complain that they are tired of their out-of-the-box agent harness producing verbose outputs. If you built a harness, that's easier to fix with a system prompt optimized for you. 

This is why Pi is getting insane adoption these days. You want to control your harness. I agree that it's hard to optimize where things stand, but self-improving algorithms will make it easier to learn from trajectories. And if you use local models, you can tune all of that end-to-end (through a co-evolution of the harness and the models) as you work with them.

Look into open-source projects like Eve, Exo, and Prime Agent, and you will see what I mean.

Uncle Bob Martin@unclebobmartin

OK.  It's time to rethink this. 

I've spend the last several weeks working on a harness that tightly constrains the agents to work the way that I want them to work.  I set up all kinds of gates, and tests, and tools, and protocols, and ...

And while I was heads-down getting that to work, the agents got a LOT better.  So much so that when I came up for air, the need for my harness was obviated.  Indeed, the need for _any_ but the most liberal of harnesses may be obviated.

Just how good these things have gotten blows me away.  I have had long debates with grok and codex about the structure of systems -- as if they were senior engineers.  They often disagree with me and have their own perspectives.  I have, more than once, found myself agreeing with their views.

I have not given up on constraints and tooling.  Unit testing is still important.  So is CRAP and Mutation testing.  These tools still find bugs and offer useful constraints, though they can leave scars.  

However, the agents have gotten so good that I can now give one a very significant task with a few guidelines and it will faithfully implement it.  I can walk away for 40 minutes and when I return it will be done.  CRAP will be satisfied, Coverage will be high, and Mutation testing complete.  The architecture will be clean, and the code will be very good.

The end result may not behave perfectly, but it's so close that a couple of tweaks usually puts it into place. 

What does this mean going forward?  I'm not sure.  But I'm beginning to think that harnesses should not treat agents as components within a software design.

💬15🔄11❤️112👀12501📊37

I am extremely bullish on the Pareto frontier of collective intelligence. While everyone works on ...

elvis (@omarsar0) — Fri, 11 Sep 2026 15:36:00 GMT

I am extremely bullish on the Pareto frontier of collective intelligence. 

While everyone works on the frontier curve of single models, there is an emerging layer of compounding systems operating way above that. This is the layer you want to be operating on. Why? Well, you can swap out and swap in any model you want (closed or open). You have more control over price, performance, models to use, and how much of that intelligence stack you want to own.

hardmaru@hardmaru

Introducing Fugu Max and Fugu Ultra v2: Orchestrating the Pareto Frontier

The AI industry has spent a decade optimizing along a single axis: build bigger, more expensive models. But intelligence has never been a monolith. It is a collective, distributed system. Humanity itself is a collective intelligence.

We built Sakana Fugu on this conviction: the most powerful AI systems will not be isolated giants, but collaborative ecosystems that learn to coordinate. Evolution innovates under constraints, and the future belongs to systems that know not just how to solve a problem, but which machinery to deploy for the lowest possible cost.

Today we are releasing Fugu Max and Fugu Ultra v2.

Fugu Max expands the Pareto frontier outward, orchestrating our largest pool of open and specialized models to deliver frontier-grade results at a fraction of the token spend. Fugu Ultra v2 pushes that frontier upward, achieving peak performance on complex multi-step tasks without depending on the very frontier models it competes against.

Together, they show that orchestration does not force a choice between cost and performance. It can push both at the same time, advancing the Pareto frontier.

Relying on a single company’s model for critical infrastructure is a massive risk. As recent export controls have shown, access can disappear overnight. Collective intelligence is the practical hedge against this concentration of power. An orchestration system simply routes around vendor restrictions by relying on an entirely swappable agent pool.

I am incredibly proud of our team for shipping this. By orchestrating the world’s models, we are building the resilient infrastructure required for AI sovereignty.

💬8🔄3❤️23👀9122📊9

Harness engineering is a top skill right now This new Meta paper is a good production example. Aut...

elvis (@omarsar0) — Fri, 11 Sep 2026 15:01:05 GMT

Harness engineering is a top skill right now

This new Meta paper is a good production example.

Auto-RecSys runs autonomous research on Meta's industry-scale recommendation models, where one training run can take days.

It runs experiments in parallel across servers, keeps a shared memory so work survives failures and new sessions, and splits guidance into natural-language skill files for reasoning and deterministic scripts for anything operational.

Two loops improve it over time.

Model-specific playbooks record failed attempts and keep working pipelines. Experimental results feed the next round of ideas.

As the playbook matured, major fixes per iteration fell from 4.0 to 1.3, and the failures fell into repeatable categories.

Paper: arxiv.org/abs/2609.10922

Chat with Paper: academy.dair.ai/papers/auto-re…

💬18🔄35❤️219👀10971📊78

This is a brilliant paper. It's of the cleanest long-context agent designs I have seen in the past ...

elvis (@omarsar0) — Thu, 10 Sep 2026 20:05:02 GMT

This is a brilliant paper.

It's of the cleanest long-context agent designs I have seen in the past couple of months.

Sequential memory agents read chunks one after another while maintaining a compact memory state.

This behavior ties reasoning depth to document traversal and makes accuracy sensitive to where the evidence sits. It also makes latency grow linearly with document length.

PARSER decouples the two.

A bank of lightweight subagents, each bound to a single chunk, reads the whole document in parallel.

A lead agent reasons through iterative scatter-gather rounds, broadcasting a query to all subagents, aggregating the returned evidence, and forming a deeper follow-up query conditioned on what it has found.

All the learnable behavior is build into the lead agent, which is trained with RL. The subagents stay frozen off-the-shelf models.

On multi-hop QA from 7K to 896K tokens, a 4B PARSER beats the strongest sequential memory baseline by 5.7 points on average and 12.0 points at 896K. At 9B it passes DeepSeek-V4-Pro by 6.3 points.

Controlled experiments show it holds up under perturbations to evidence position, order and distance, which cause large accuracy swings in sequential methods, while cutting inference latency by up to 11x.

Paper: academy.dair.ai/papers/parser-…

💬18🔄37❤️195👀13796📊74

Interesting read on how Pocket FM scaled. Pocket FM makes serialized audio fiction. AI now writes a...

elvis (@omarsar0) — Thu, 10 Sep 2026 18:13:56 GMT

Interesting read on how Pocket FM scaled.

Pocket FM makes serialized audio fiction. AI now writes and voices much of it, taking annual output past 2.5M hours.

Daily listening time went from ~25 mins to >150 mins.

More than 500,000 people now write on the platform, and a lot of them started as listeners.

Rohan Nayak@RohanNayak2

We grew from ~0 to $500M ARR, adding $250M last year alone while being EBITDA profitable.

1,300-word post on every growth tactic that worked for us:

1. What got us from 0-$400M ARR in the US works in every country.
2. Retention and Monetisation Hacks.
3. Localization > Translation.
4. Experiment with $1M internal seed checks.
5. Pivoting to only AI content production unlocked 100% ARR growth.

Pocket FM is like Netflix for audio-only dramas, with our own pool of one-person studios.

1. Acquisition Playbook for 0-$10M ARR in any country in 6 months

We run a 90 sec video trailer of an audio drama as an Ad and ask users to download the app if they're interested in the rest of the story.

Our core insight after spending >$100M on acquisition is: If the clickthrough rate (CTR) for an ad goes from 2% to 2.25%, our customer acquisition cost (CAC) decreases by ~ 30%.

We remodelled our system around this insight and built an AI-first 2.25% CTR ad manufacturing machine that works in every country.

For every new country, we take our hit shows -> use LLMs to extract and most intriguing moments -> Write a 5-min script combining all the best parts. The first 60 seconds has to have a hook every 5 seconds and needs to end with a crazy cliffhanger to force a download mid-scroll.

If a marketing video is not hitting our benchmarks (2.5% CTR and 55% 3-second through-play), a creative director gets involved to change the hook or cliffhanger to get the numbers there.

AI lets us make 1,000 Ads per show, and in total we do ~17.5k Ads per month. When a business creating scales from 1k to 10k Ads, the normal thing is for CAC to skyrocket. But with what I just shared, we 7-8x'd our User Acquisition budget without increasing our CAC materially.

It took us 8 months to figure this out, but then the timeline from 0-$10M in every country got shorter and shorter:

US revenue grew to $25M in 19 months. (US is now 78% of total)
Germany to 21M in 11 months.
France to $10M in 3 months.

2. Retention and Monetisation Hacks.

We knew we wanted to create an audio entertainment platform, but there was no standard format. For the first 2-3 years, we tried 10 different formats before landing on the winner.

After we got it right with audio drama (8-12min chapters, written for mobile fiction, serialized, episodes have strong hooks and end on cliffhangers), avg daily streaming time went from ~25mins to 150+.

Audio drama made us realize that Pocket FM was creating a whole new medium. There was no playbook for anything that we were doing. Everything had to be thought & built from scratch.
This is what we did for each major bullet:

- Monetization: Users have some free daily minutes to listen; then it's pay per episode. We also added the option for users to unlock episodes by watching ads, which is doing extremely well. Ads scaled from zero to a ~$90M run rate in 12 months.
- Engagement & production: You don't become obsessed with an app you open once a week. To have users engage daily, we make the next episode free every day. Also helps them build the habit.
- Discovery. Huge problem because people consuming Pocket FM enter the app, tap the show, and lock the screen. We fixed it by doing "playlists" of episodes. Once users's free minutes on a show are done, we ask users if they want to pay. If they say no, we play a new show, one where they haven't used their free mins. And we stitch episodes of different shows together that way, creating natural discovery.

3. Localization > translation.

78% of our revenue is still concentrated in the US. Localisation is fixing this: You might write a show for a Spanish audience where language, jokes, folklore, have a certain flavor; if you merely translate the show for, say, a Norwegian audience, that color is lost and hurts the show in the Norway. Listeners would relate more if the show was written by a Norwegian.

That's why, instead of translating, we localize shows to different regions. We use AI to take the spine of stories and adapt their whole cultural layer to the other country.

The results: a US show that was localized for a German audience had 50% higher retention than the translated version.

Localization + our user acquisition playbook led to:
- $10M+ ARR in France within 3 months
- $21M+ ARR in Germany within 11 months.

And this expands writers' addressable market.
We get messages of writers thrilled to have revenue coming from the US, India, EU, LatAm, without them doing much incremental work.

4. Experiment.

We give $1M checks to new internal initiatives, and the team has 12-18 months to prove their thesis. If they prove it, we double down.
This is how Pocket Saga came about, our AI video microdrama app. It's an AI video equivalent of Pocket FM. Same shows and structure, but as a vertical 2-min mobile video series.

We launched it two months ago and it's at ~$15M ARR.

Pocket's broader thesis is to help creators tell their stories to as many people as possible. Start with audio drama -> multiple languages or localize to diff countries -> microdrama -> more formats like movies, TV shows, and games.

Best of all is that writers get revenue streams not only from countries they wouldn't have tapped into, but also from formats they wouldn't have thought possible.

5. Pivoting to only AI content production unlocked 100% ARR growth.

In mid-2024, we pivoted to only AI content production.
Our growth took a very direct hit. Pocket was already at $200M ARR, growing 50% YoY, and we flatlined for a whole semester.

Six months later, the business exploded.

After the switch, we went from ~25k hours of content produced per year to over 2.5M hours, which are also higher quality. Because AI orchestrates the writing and more data is fed into it, more blockbusters come out.
Over 90 titles have $1M+ in lifetime earnings and 13 crossed $10M.

Quality control is done with LLM as a judge, and LLMs are quite tough.
Results are very encouraging:
- 12-month revenue retention went from 44% to 76%.
- In the past year, we added $250M in net new ARR.
__________________________________

Pocket Entertainment (Pocket FM + Pocket Saga) has become the largest AI entertainment platform.

We have the largest storytelling catalog with 770,000 titles, 550k creators, and 5.5B hours of playtime with minute by minute retention & engagement data.  We are using all of this data to improve every aspect of our business.

Our Bet: In the next 3 years we will be able to produce Naruto-like series for $1000.

Netflix has to spend $17B to find 100 Blockbusters per year.
What happens when you can produce 1M high quality shows for $1B?

The streaming wars caused a $300B reallocation of market cap.
The AI entertainment wars may be a $1T+ reshuffling of market cap.

We are at a unique spot. Unlike other AI creator tools like Runway or Midjourney, we own both the supply and demand side of AI content.

We've attracted over 550,000 writers who are producing an annualized 2.5M hours of content every year. This pairs with 137B+ minutes streamed.

Because of this, growth is accelerating as we scale further.
We have multiple S-curves inflecting at the same time:

- AI produces better and more ads -> scale faster in new countries
- ⁠Writers + AI produce more shows -> more blockbusters
- ⁠Localisation -> more reach per blockbuster
- ⁠Audio to video format expansion -> larger TAM -> more creators

Every aspect of our business is designed to improve another, and everything is aligned so that the number of blockbusters continues to rise.

If you're interested in building at the intersection of ai, tech and entertainment, DM me.

💬3🔄2❤️15👀10106📊5

Just amazed by how specialized models keep pushing the Pareto frontier. Just watch that massive co...

elvis (@omarsar0) — Thu, 10 Sep 2026 16:00:30 GMT

Just amazed by how specialized models keep pushing the Pareto frontier. 

Just watch that massive cost reduction.

SWE-2 achieves 50.0% on FrontierCode 1.1 Main1, within one point of Fable 5.1, while being 64% cheaper.

Cognition@cognition

Introducing SWE-2, our closest model yet to the frontier.

On leading evals, it scores on par with recent frontier models – at up to 70% lower cost.

We scaled RL to multiple trillions of parameters, with a refined recipe that pushes the Pareto curve on both capabilities & cost.

💬14🔄8❤️79👀10256📊19

This is brilliant—a great program to get started with building agents. Lots of really great resour...

elvis (@omarsar0) — Thu, 10 Sep 2026 15:49:17 GMT

This is brilliant—a great program to get started with building agents. 

Lots of really great resources to learn how to build and scale with AI.

Includes open models, orchestration, retrieval, evaluation, and observability on top of it.

Nebius@nebiusai

The Nebius AI Builder Program is now available.
AI isn’t just a model you call anymore. It’s a system you build. And builders, not a few closed labs, will decide what it becomes.

The open ecosystem has all the pieces. We want to make it easier to put them together and start building.

The program is free, with $400+ in credits and discounts, working code and cookbooks, office hours with engineers, and a community to build with.

We’re joined by @NVIDIAAI, @LangChain, @huggingface, @cognition, @OpenHandsDev, @tavilyai, @TolokaAI, @composio, @PrimeIntellect, @MiniMax_AI, @Alibaba_Qwen, and more joining soon.

Join with the link in the comments 👇

Your browser does not support the video tag.

💬4🔄6❤️39👀10294📊9

Great paper from AWS. I use a similar setup where an agent orchestrator sits on top of a multi-agen...

elvis (@omarsar0) — Thu, 10 Sep 2026 15:30:16 GMT

Great paper from AWS.

I use a similar setup where an agent orchestrator sits on top of a multi-agent system.

(bookmark it)

This work introduces one of the many approaches available to manage compound LLM systems.

Compound LLM systems usually solve coordination by adding a higher-level model.

That meta-agent reads worker outputs, writes the final answer, allocates later calls and decides when to stop, which concentrates three separate control decisions in one opaque, order-sensitive call.

UnitBoost investigates whether the manager needs to be generative at all.

A task-given unit map turns worker outputs into slot-value proposals, a constrained argmax assembles the output, and slots left unfilled or unsupported become an explicit residual that directs the next round.

On three held-out benchmarks it beats the best single candidate chosen with gold labels by 0.060 to 0.195 task-score points, and beats input-matched generative managers by 0.048 to 0.076.

Replacing only the management step improves six compound-system configurations. Residual-directed rounds raise FanOutQA cell F1 from 0.4778 to 0.5524.

Chat with Paper: academy.dair.ai/papers/unitboo…

💬13🔄10❤️89👀9911📊37

Another goated release by DeepSeek. Open weight, btw. Outperforms Opus 5 on key benchmarks, btw. L...

elvis (@omarsar0) — Thu, 10 Sep 2026 15:02:03 GMT

Another goated release by DeepSeek. 

Open weight, btw. Outperforms Opus 5 on key benchmarks, btw. Lower prices too. 

Looks like an extremely efficient model. 

What a legendary run.

DeepSeek@deepseek_ai

🚀 Introducing DeepSeek-V4.1-Flash: smarter, faster, more efficient.

🔹 Introducing the smallest model in our new architecture family, with native visual understanding.
🔹 Designed for greater capability, faster inference, higher throughput, and scaling to larger models.

1/6

💬4🔄7❤️47👀9280📊9

Nice paper from Salesforce on co-evolving harnesses and models. Harness engineering is a hot topic ...

elvis (@omarsar0) — Thu, 10 Sep 2026 08:00:08 GMT

Nice paper from Salesforce on co-evolving harnesses and models.

Harness engineering is a hot topic right now. So this is a great read.

(bookmark it)

Salesforce evolved a harness with a weak model across seven enterprise agent tasks, then trained that model on a stronger expert's full trajectories under the same harness.

Performance dropped on all seven tasks, by 4 to 30 points across Qwen3-Coder and Gemma 4.

The same fine-tuning helps under the unevolved harness. So the harness is what changes the outcome.

Their analysis points at model-harness fit.

Imitation transfers knowledge and increases scaffold usage, but the weaker model adopts the expert's planning strategy without the competence to execute it, and it no longer matches a harness that was evolved around its own native planning style.

The fix is to stop copying whole trajectories.

A meta-level agent finds the failing turn in the weaker model's own rollout and asks the expert to rewrite only that turn. That keeps the model's planning style intact and combines the gains from harness evolution and weight updates.

Paper: academy.dair.ai/papers/co-evol…

💬23🔄23❤️119👀10985📊55

This is a really neat benchmark. It's super interesting that Astra starts strong and holds the lead ...

elvis (@omarsar0) — Wed, 09 Sep 2026 23:02:04 GMT

This is a really neat benchmark. It's super interesting that Astra starts strong and holds the lead for up to 19 hours, but Fable 5.1 catches up in the final hours.

Not entirely clear why.

Different models solve long-horizon problems with different strategies. It's no surprise that we see different trends. 

And then there is also the question of cost-performance efficiency. Qwen3.8 Max, Gemini 3.8 Flash, and Grok 4.6 are more economical for research, earning a place on the cost–performance Pareto frontier.

For AI research, things get complicated not because of the duration of the task but because agents (even with the best models) tend to stagnate due to low-quality exploration. Might be due to OOD. Or simply that agents are just simply not "creative" enough (well, at least not at human levels), which the top researchers are exceptional at, from years of experience building intuition and deep expertise. 

The other question is: where exactly are research agents spending their compute? Are they using it efficiently? Earlier work from @intology found that coding agents mostly spend compute on hyperparameter tuning, rarely attempting algorithmic research. x.com/intology/statu… Not sure where the research stands on that today, but all of these are interesting research questions.

Alex Dimakis@AlexGDimakis

We are releasing AutoResearchExam, a benchmark on open-ended machine learning and engineering tasks. Our benchmark covers seven research areas including model training, data curation, AI safety and interpretability.

benchmarks.bespokelabs.ai/autoresearchex…

In each task, we give agents 24 hours with a CPU or GPU machine to develop and improve their solutions through experiments and feedback. We measure both speed and quality with a combined score. Our benchmark has a unique feature: testing if agents create improvements that hold up on data they never see. We find that AI research agents often overfit as they try to improve. 

We see an interesting head-to-head comparison at the frontier: Astra starts the strongest and holds the lead for up to 19 hours but Fable 5.1 catches up and gets the top performing spot in the final hours. 

Qwen3.8 Max, Gemini 3.8 Flash and Grok 4.6 all sit on the cost-performance Pareto frontier, giving strong options at lower API budgets.

Anthropic's Opus and Fable retain nearly all their validation performance on hidden tests, with gaps of 1.1% and 2.9%. Astra's improvement over Sol extends to generalization too, with that gap falling from 6.9% to 1.7%. (1/n)

Your browser does not support the video tag.

💬9🔄6❤️44👀11667📊13

Highly recommended. Model-harness co-optimization is where you design the harness for the task and t...

elvis (@omarsar0) — Wed, 09 Sep 2026 20:55:09 GMT

Highly recommended. Model-harness co-optimization is where you design the harness for the task and then tune the model inside the harness.

A clear example of why you want to own the harness and the models. Harness engineering is the real deal. So much to unlock.

I have been tracking a bunch of papers showing how extremely effective this is. And there are many variants. Harvey explores RLMs, which somehow enable interesting ways to scale things. 

A few links here already: academy.dair.ai/papers/collect…

Niko@nikogrupen

x.com/i/article/2097…

💬14🔄18❤️171👀21373📊49

Another banger paper from Google. If you build memory for long-horizon agents, this one is worth yo...

elvis (@omarsar0) — Wed, 09 Sep 2026 18:34:02 GMT

Another banger paper from Google.

If you build memory for long-horizon agents, this one is worth your time.

(bookmark it)

Really nice to see how knowledge graphs are being explored in creative ways for agents. This has lots of implications for self-evolving agents.

Technical summary below:

Agents usually pick actions by generating over an accumulating history, which leaves the procedural knowledge implicit. As trajectories get longer they lose track of objectives, call tools out of order, and repeat actions that did not work.

The Procedural Graph helps to make that knowledge explicit.

A knowledge graph stores facts as entity-relation-entity triplets. A Procedural Graph stores procedures as procedure-relation-procedure triplets, so the agent can query what to do next and under which conditions.

At each step, the framework localizes the agent's active node, and a guidance model turns the surrounding subgraph into step-level guidance that biases the next action without dictating it.

The graph rewrites itself.

An LLM refiner compares failed trajectories against successful ones and edits the topology and attributes, committing only edits that hold up on held-out validation, and keeping the rejected ones on file so the same change is not proposed twice.

Starting from a minimal skeleton it builds graphs that match or beat hand-designed ones, and it helps to repair a flawed expert priors instead of inheriting them.

Paper: academy.dair.ai/papers/procedu…

💬46🔄141❤️1028👀87669📊453

Better models don’t automatically mean smaller AI bills. Sometimes they mean you finally have a rea...

elvis (@omarsar0) — Wed, 09 Sep 2026 18:24:10 GMT

Better models don’t automatically mean smaller AI bills.

Sometimes they mean you finally have a reason to run the workflow you couldn’t get working before.  

That’s what makes GPT-6 Astra interesting to me.  

Once a model can research, build, inspect, and refine, you start giving it more ambitious tasks. A single response becomes an entire working session.  

The useful question is no longer just “what does a million tokens cost?”  

It’s “what can I get done with this budget?”  

That’s also why cashback on inference is worth paying attention to. On work you were already going to run, getting some of the spend back gives you a choice: keep the savings or fund another experiment.  

But the goal should never be to burn more tokens.  

It should be to get more verified, useful work out of the same budget.

Mujtaba@Mutchtaba2

STOP PAYING FOR TOKENS !

START GETTING PAID INSTEAD

Introducing Straitly, the first AI service that PAYS YOU for using tokens

Unlike
> Openrouter charges 5.5% service fee
> Vercel charges 2.9% payment fee
> Anthropic gives you low rate limits
> OpenAI only lets you use GPT models

Straitly:
> PAYS YOU 5% CASHBACK on all closed-source models
> 10% CASHBACK on open-source models
> High rate limits
> 177 AI models
> 99.8% reliability
> ZDR

This means that you can save up to 25% on your token bills compared to using services like OpenRouter or Vercel, while still getting the same reliability and model choices

Since our waitlist launch 3 weeks ago we have processed ~150B tokens already, and are growing at 35% week over week

We're backed by @_LeapYear_, Sand Hill North, pif.co and other incredible angels in our mission to build the utility layer for all future token usage

This is just the beginning

Today Straitly is live and open for public access

Start getting paid today at straitly.ai

To celebrate, we are giving out $10k in free AI credits to 400 people

Comment below what your are building and what model you use the most, then head over to Straitly and sign up

We'll select the winners at the end of September !

Your browser does not support the video tag.

💬3🔄1❤️20👀10214📊5

Better models don’t automatically mean smaller AI bills. Sometimes they mean you finally have a rea...

elvis (@omarsar0) — Wed, 09 Sep 2026 17:34:08 GMT

Better models don’t automatically mean smaller AI bills.

Sometimes they mean you finally have a reason to run the workflow you couldn’t get working before.

That’s what makes GPT-6 Astra interesting to me.

Once a model can research, build, inspect, and refine, you start giving it more ambitious tasks. A single response becomes an entire working session.

The useful question is no longer just “what does a million tokens cost?”

It’s “what can I get done with this budget?”

That’s also why cashback on inference is worth paying attention to. On work you were already going to run, getting some of the spend back gives you a choice: keep the savings or fund another experiment.

But the goal should never be to burn more tokens.

It should be to get more verified, useful work out of the same budget.

💬0🔄0❤️2👀1499📊1

It's crazy how far small models can be pushed. Robbyant just open-sourced LingBot-World 2.0 Small, ...

elvis (@omarsar0) — Wed, 09 Sep 2026 16:53:28 GMT

It's crazy how far small models can be pushed.

Robbyant just open-sourced LingBot-World 2.0 Small, a 1.3B world model that generates an interactive world in real time on a single consumer GPU.

Same pattern we saw with language & image models. Scale capabilities first, then distill down to hardware people own.

Robbyant@robbyant_brain

🌍 What if an AI-generated world didn’t end after a few seconds—but kept running for hours, responded to every move, let you attack, cast spells, shoot, or summon storms, and continued evolving through AI agents?

That’s LingBot-World 2.0: our open-source real-time interactive world model, scaling to 720p/60fps in our full setup.

Today, we’re open-sourcing 3 more models:
🔹 LingBot-World 2.0 Small (1.3B) 
🔹 LingBot-World 2.0 Bidirectional  
🔹 LingBot-World 2.0 Causal Pretrain 

And now, with our 1.3B Small Model, that world can run in real time on a single consumer GPU. 🧵

Your browser does not support the video tag.

💬6🔄9❤️95👀13063📊22

We should all be building for agents. @lightfld validates this with an impressive CRM solution for ...

elvis (@omarsar0) — Wed, 09 Sep 2026 16:21:26 GMT

We should all be building for agents.

@lightfld validates this with an impressive CRM solution for the agentic era. And massive adoption, too.

It gives you API, MCP, and CLI access to every record so automations can easily be built on top of it. Removes all the complications.

Lightfield@lightfld

We’ve raised $47M in Series A funding led by @a16z to reimagine CRM as a world model of a business.

💬8🔄1❤️20👀9882📊8

See more here: https://t.co/Aszx9ny7yU

elvis (@omarsar0) — Wed, 09 Sep 2026 14:55:45 GMT

See more here: huggingface.co/Lightricks/LTX…

💬0🔄0❤️4👀1252📊1

Open video models are having their moment. The previous generation of LTX alone reached 18M downloa...

elvis (@omarsar0) — Wed, 09 Sep 2026 14:55:44 GMT

Open video models are having their moment.

The previous generation of LTX alone reached 18M downloads, which shows how much demand there is for video models that builders can own and modify themselves.

LTX-2.5 from @ltx_io builds on that foundation as a world model with open weights, local deployment, and a pretrained base that teams can fine-tune for their own work.

For me, it's all about owning the intelligence stack.

Builders choose the hardware, keep access to the weights, and can fine-tune the model around a creative or production workflow.

More capabilities related to this model:

The release improves both generation and editing.

The new decoder targets sharper faces, legible text and signage, and cleaner fast motion.

Native Multishot generates connected shots while preserving character, environment, lighting, and voice across cuts.

IC-LoRA works on footage you already have, with support for object removal, continuity and wardrobe fixes, and environment changes without a reshoot or frame-by-frame rotoscoping.

A stronger distilled model brings more of the full model's quality and motion to local GPUs.

Open video is still early, and releases like this give builders more room to experiment, specialize models, and own the production stack.

I am looking forward to seeing what people build with LTX-2.5.

Your browser does not support the video tag.

💬12🔄5❤️31👀8864📊18

Super cool paper from Microsoft. They show that it's possible to build competitive small coding ag...

elvis (@omarsar0) — Wed, 09 Sep 2026 14:40:52 GMT

Super cool paper from Microsoft. 

They show that it's possible to build competitive small coding agents without traditional distillation from frontier models.

DAIR.AI@dair_ai

Banger report from Microsoft.

(bookmark it)

They show that it's possible to build competitive small coding agents without traditional distillation from frontier models.

This is a big deal!

The work describes how they achieved this.

They introduce a 4B coding agent trained on roughly 1,500 software engineering environments.

The cool thing is that they use no distillation from a larger model at any point.

FrogNano is post-trained purely with RL on synthetic tasks.

The target is a coding agent that runs on minimal machines, which rules out both a frontier backbone and a frontier teacher.

The ingredient the report credits the most is online task synthesis.

The pipeline generates tasks calibrated to the frontier of learnability for the current checkpoint, so the agent always trains on problems it can just barely solve. The authors argue that calibration, rather than the volume of synthetic data, is what makes this work.

This means that competitive small coding agents can be trained from synthetic tasks alone.

And generating those tasks at the current agent's learnability frontier is what makes this particular training productive.

The report covers training methodology, evaluations across diverse environments, and analyses of what the agent learned.

Paper: academy.dair.ai/papers/frognan…

💬13🔄17❤️134👀19556📊36

Human review and the PR gate stay in place. This adds a quality layer before code ever gets there. ...

elvis (@omarsar0) — Wed, 09 Sep 2026 13:18:06 GMT

Human review and the PR gate stay in place. This adds a quality layer before code ever gets there.

Install it with one command in the terminal, or from the Claude Code, Codex, or Kiro marketplace. It also runs as an MCP server for any MCP-compatible agent.

More here
qodo.ai/features/qodo-…

Thanks to the Qodo team for partnering on this post.

💬1🔄1❤️5👀1138📊2

The findings were specific. My new ingestion path marked papers as public without checking for full...

elvis (@omarsar0) — Wed, 09 Sep 2026 13:18:05 GMT

The findings were specific.

My new ingestion path marked papers as public without checking for full text. That is the exact check the refactor was meant to add.

The chat endpoint was accepting questions on papers with no verified text.

Summaries were being stored as full documents.

A paid lesson page was missing an enrollment check.

All of this surfaced before a PR existed.

💬1🔄1❤️4👀1218📊2

Qodo just launched the Agentic Toolbox. It gives Claude Code, Codex, Kiro, and Cursor direct access...

elvis (@omarsar0) — Wed, 09 Sep 2026 13:18:02 GMT

Qodo just launched the Agentic Toolbox.

It gives Claude Code, Codex, Kiro, and Cursor direct access to @QodoAI's review engine, codebase knowledge, and team rules.

💬8🔄4❤️28👀9029📊14

Nice paper to improve inference efficiency. It's been a while we haven't seen good work on efficien...

elvis (@omarsar0) — Wed, 09 Sep 2026 05:33:02 GMT

Nice paper to improve inference efficiency.

It's been a while we haven't seen good work on efficiency.

Here is why it matters:

A long-running agent's workspace outgrows its context window long before the task finishes.

The first approach commonly used, compaction, loses the fine-grained execution evidence. And text retrieval re-prefills content the model already processed.

KVMem keeps the overflow as paged KV state instead, spread across GPU memory, host memory and NVMe.

Lightweight attention-space indexes, native to the model, pick the relevant historical blocks and materialize a query-dependent view that fits inside the native context window.

On the DeepSWE long-context test with Qwen3.8-27B, task success goes from 43.8% under compaction to 48.4%.

The local deployment result stands out. It runs Qwen3.6/3.8-27B NVFP4 with MTP on a laptop with a 24GB RTX 5090, virtualizing an agent workspace up to 1M tokens, four times the model's native 256K window, at around 50 tokens per second.

Paper: academy.dair.ai/papers/kvmem-v…

💬23🔄8❤️60👀9018📊37

I highly recommend the list of papers, as they cover some of the most impressive ideas our harnesses...

elvis (@omarsar0) — Tue, 08 Sep 2026 22:28:42 GMT

I highly recommend the list of papers, as they cover some of the most impressive ideas our harnesses in the past couple of years. 

You can chat with the papers too. 

Here is the YC-related talk: youtu.be/n9xKblqyQ28?si…

💬0🔄3❤️19👀4240📊8

Harness engineering is a top skill right now. I saw a great overview of harness engineering in a re...

elvis (@omarsar0) — Tue, 08 Sep 2026 22:16:56 GMT

Harness engineering is a top skill right now.

I saw a great overview of harness engineering in a recent YC Paper Club talk.

It's a really good list of top papers.

So I decided to build it as a collection. Enjoy!

💬23🔄92❤️795👀50449📊307

Find the whole collection here: https://t.co/hSkMMHjf1L

elvis (@omarsar0) — Tue, 08 Sep 2026 22:16:56 GMT

Find the whole collection here: academy.dair.ai/papers/collect…

💬2🔄8❤️37👀5115📊21

Great to see Meta getting into always-on proactive agents. Given their distribution, it's going to...

elvis (@omarsar0) — Tue, 08 Sep 2026 19:44:47 GMT

Great to see Meta getting into always-on proactive agents. 

Given their distribution, it's going to be interesting to see how ready people are for personal agents.

Fig: Muse under the hood.

Mark Zuckerberg@finkd

Introducing Muse, the personal agent that understands your goals and works 24/7 to get things done for you.

💬10🔄7❤️82👀12635📊21

Recent advancements in math with the use of AI propel us into an undeniably exciting new arc of AI-e...

elvis (@omarsar0) — Tue, 08 Sep 2026 18:37:59 GMT

Recent advancements in math with the use of AI propel us into an undeniably exciting new arc of AI-enabled scientific discovery and understanding. And yet, it feels like we are just scratching the surface. Stay humble, stay curious, and stay focused on what matters most.

💬6🔄5❤️29👀8701📊9

🤯 For the Navier-Stokes Millennium Prize Problem solution, OpenAI says they used a next-generation ...

elvis (@omarsar0) — Tue, 08 Sep 2026 17:31:10 GMT

🤯 For the Navier-Stokes Millennium Prize Problem solution, OpenAI says they used a next-generation model that is significantly more capable than GPT-6 Astra.

If GPT-6 Astra is this good already, imagine what the next generation will be like.

OpenAI@OpenAI

We’re sharing a solution to the Navier-Stokes Millennium Prize Problem, one of the deepest problems at the frontier of mathematics.

The proof was produced by a group of agents, using an OpenAI next-generation model significantly more capable than GPT-6 Astra.

The problem concerns whether the description of smooth three-dimensional fluid motion modeled by the Navier-Stokes equations can break down. It has remained unresolved for roughly 90 years.

💬9🔄5❤️61👀12388📊12

Nice report on agents beyond code generation. Here is why it matters: Coding agents raise how much...

elvis (@omarsar0) — Tue, 08 Sep 2026 16:55:02 GMT

Nice report on agents beyond code generation.

Here is why it matters:

Coding agents raise how much code gets written. This report argues the gains shrink sharply between writing code and shipping reliable software.

It pulls together field studies, benchmark audits and production reports from 2024 through September 2026. What stays constraining is review, integration, testing, security, deployment and production operations.

The cost side changes shape too.

Predictable per seat licensing gives way to variable token, tool, sandbox, CI and rework costs, which is a different budgeting problem than buying licenses.

Four interesting concepts emerged in this report. The Agentic SDLC Throughput Paradox, Production-Qualified Change, the Verification Tax, and an Agentic SDLC Control Plane that allocates autonomy under explicit cost, reliability and human attention budgets.

Paper: academy.dair.ai/papers/beyond-…

💬22🔄18❤️88👀9794📊45

Let me know if you have any questions as I prepare the resources, prompts, etc., to help others buil...

elvis (@omarsar0) — Tue, 08 Sep 2026 16:00:37 GMT

Let me know if you have any questions as I prepare the resources, prompts, etc., to help others build their own.

💬3🔄0❤️5👀1337📊3

It feels like GPT-6 Astra has no limits. It just built this human anatomy with ~4K structures. I w...

elvis (@omarsar0) — Tue, 08 Sep 2026 15:57:29 GMT

It feels like GPT-6 Astra has no limits.

It just built this human anatomy with ~4K structures.

I wanted to look closely at what Astra was doing.

Notes:

I can explore at different levels and select any body part for closer inspection with unique views. Amazing!

If you give Astra instructions to use any tool or resource, it does some mind-blowing things. 

Make sure when you prompt it to not hold back or constrain it too much. Allow it to search and explore.

This is the most excited I have been about AI. This opens up personalized learning, understanding, and applications unlike anything possible before. 

To be clear, it's not about originality. The model used existing works and resources. But even so, it's probably the best model at doing research, which is needed to get these working prototypes. 

The personalization of building apps like this to your needs- that is the most exciting part of the Astra demos you have been seeing in your timeline. That level of personalization unlocks unique potential and creativity in all of us. IMO, the more accurate term describing new models like Astra is "personalized general intelligence." 

Stack the model used (self-selected): React 19 and TypeScript, Vite for builds, Three.js with OrbitControls for interactive 3D, plain CSS, Lucide icons, and Vitest with React Testing Library. The anatomy models come from Z-Anatomy/BodyParts3D.

Will share prompts and process soon. Stay tuned.

Your browser does not support the video tag.

💬12🔄12❤️142👀12287📊41

I also wrote a little interactive tutorial if you want to test it yourself. I highly recommend that...

elvis (@omarsar0) — Mon, 07 Sep 2026 20:45:02 GMT

I also wrote a little interactive tutorial if you want to test it yourself.

I highly recommend that you try it at least once and pay close attention to the output improvement. 

academy.dair.ai/resources/writ…

💬2🔄6❤️101👀22044📊48

A few days ago, Anthropic shared this brilliant prompt. I was surprised by how it significantly im...

elvis (@omarsar0) — Mon, 07 Sep 2026 20:29:04 GMT

A few days ago, Anthropic shared this brilliant prompt. 

I was surprised by how it significantly improved writing with Fable 5.1.

Even more surprising, it also improved writing for GPT-5.6 Sol. 

I have it as a rule everywhere I use AI for editing/writing.

💬122🔄219❤️4275👀292792📊1892

GPT-6 Astra takes about 30-60 minutes to generate these. It has nice verification steps that it uses...

elvis (@omarsar0) — Mon, 07 Sep 2026 18:39:46 GMT

GPT-6 Astra takes about 30-60 minutes to generate these. It has nice verification steps that it uses to make sure things don't look off. I think most people won't understand how insanely difficult this is to build from scratch. AI models just weren't able to do this well in my previous test. GPT-6 Astra is the first model that goes over the threshold I was looking for.

💬1🔄0❤️3👀1590📊1

With everything put together, it feels like less slop and something I would sit and watch.

elvis (@omarsar0) — Mon, 07 Sep 2026 18:33:38 GMT

With everything put together, it feels like less slop and something I would sit and watch.

💬2🔄1❤️5👀2107📊3

OMG!!! Using GPT-6 Astra to animate fundamental concepts. It produced this entire clip. I wish I h...

elvis (@omarsar0) — Mon, 07 Sep 2026 18:29:14 GMT

OMG!!!

Using GPT-6 Astra to animate fundamental concepts. It produced this entire clip.

I wish I had GPT-6 Astra when I was learning this stuff.

This really changes personalized learning.

Your browser does not support the video tag.

elvis@omarsar0

This is insane! 

GPT-6 Astra built this beautiful math animation in one go! (🔉 sound on)

"Jaw-on-the-floor" moment. 

I've not been able to get anything close to this with previous AI models. 

True personalized learning has arrived!

Your browser does not support the video tag.

💬14🔄10❤️186👀35300📊43

View it here: https://t.co/4oQLNAjWhC Let me know what you think. I will continue to explore this m...

elvis (@omarsar0) — Mon, 07 Sep 2026 17:08:38 GMT

View it here: academy.dair.ai/resources/atte…

Let me know what you think. I will continue to explore this more in the coming days.

💬0🔄3❤️18👀2269📊6

Testing GPT-6 Astra on visualizing popular AI papers. Started with this beautifully generated inte...

elvis (@omarsar0) — Mon, 07 Sep 2026 17:08:37 GMT

Testing GPT-6 Astra on visualizing popular AI papers. 

Started with this beautifully generated interactive version of the Transformer paper. 

This unlocks new ways to explore research.

Your browser does not support the video tag.

💬17🔄21❤️193👀17675📊59

// Design Docs Are All You Need // Banger paper from Google DeepMind, MIT, and colleagues. What a ...

elvis (@omarsar0) — Mon, 07 Sep 2026 15:25:02 GMT

// Design Docs Are All You Need //

Banger paper from Google DeepMind, MIT, and colleagues.

What a genuinely strange and interesting paper this one is.

Here is the setup:

They maintain a performance-modeling library whose main branch contains almost no code.

The repository is a directed graph of natural-language design docs. Coding sub-agents regenerate the entire implementation from those docs whenever a version updates.

Every human change is an edit to a doc.

The premise is that ML performance modeling invalidates its own abstractions every hardware and model generation, and coding agents are now cheap enough that regenerating a library beats patching one.

Two things make the regeneration reliable. The design docs are written around step-by-step worked examples, which act as in-context demonstrations for the generating agents. The system is also anchored on a minimal recursively defined operator IR with symbolic cost expressions in SymPy.

Regenerated implementations reproduce hand-audited reference models to round-off precision, including DeepSeek-V3 serving on a TPU pod slice.

Paper: academy.dair.ai/papers/design-…

💬30🔄44❤️297👀23675📊127

The takeaway here is that you can't be afraid to give Astra access to tools. This model feels like i...

elvis (@omarsar0) — Mon, 07 Sep 2026 15:05:01 GMT

The takeaway here is that you can't be afraid to give Astra access to tools. This model feels like it has no limits. Everything you had done before and failed, you should be trying again with it.

💬0🔄0❤️6👀2029📊1

I gave it an ElevenLabs API key, and it just figured out the rest. I am mindblowned. Scenes are al...

elvis (@omarsar0) — Mon, 07 Sep 2026 14:47:02 GMT

I gave it an ElevenLabs API key, and it just figured out the rest. I am mindblowned. 

Scenes are all customizable. It's all code. So color theme can be changed for instance. 

I will be generating more of these to see how far math animations can be pushed.

💬1🔄3❤️9👀3086📊3

This is insane! GPT-6 Astra built this beautiful math animation in one go! (🔉 sound on) "Jaw-on-...

elvis (@omarsar0) — Mon, 07 Sep 2026 14:41:23 GMT

This is insane! 

GPT-6 Astra built this beautiful math animation in one go! (🔉 sound on)

"Jaw-on-the-floor" moment. 

I've not been able to get anything close to this with previous AI models. 

True personalized learning has arrived!

Your browser does not support the video tag.

💬26🔄11❤️158👀56698📊52

// Normalized Low-Rank Adaptation (NoRA) // They propose a one-line change to LoRA that costs nothi...

elvis (@omarsar0) — Mon, 07 Sep 2026 05:45:02 GMT

// Normalized Low-Rank Adaptation (NoRA) //

They propose a one-line change to LoRA that costs nothing and improves convergence, stability and forgetting.

LoRA initializes the up-projection to zero, which means early optimization is governed almost entirely by the down-projection. This observation tells you where to regularize.

NoRA normalizes the down-projection matrices during training. The authors also show the same normalization applied once at initialization improves standard LoRA without repeating it through training, which is the cheaper of the two options.

The benefits hold across pretraining, supervised fine-tuning and reinforcement learning. Faster convergence, better final performance, more stable training, and less catastrophic forgetting.

It adds no trainable parameters and no inference-time computation, which is what makes it broadly applicable rather than another specialized LoRA variant.

Paper: academy.dair.ai/papers/normali…

💬16🔄19❤️116👀11015📊42

Jensen Huang: "AGI has arrived" Are you feeling the AGI?

elvis (@omarsar0) — Mon, 07 Sep 2026 02:07:26 GMT

Jensen Huang: "AGI has arrived"

Are you feeling the AGI?

Jensen Huang@JensenHuang

@ChaseLochmiller@OpenAI GPT-6 Astra, trained on ~100K+ NVIDIA Grace Blackwell NVLink72. From ChatGPT to o1 to Astra in 4 years.

AGI has arrived. Congratulations @OpenAI team.

400K GPUs coming online next.

💬38🔄0❤️82👀19925📊31

Super interesting paper on proactive agents from Google DeepMind. (bookmark it) Proactive assistan...

elvis (@omarsar0) — Mon, 07 Sep 2026 02:00:04 GMT

Super interesting paper on proactive agents from Google DeepMind.

(bookmark it)

Proactive assistance usually means autocomplete.

Researchers asks what it looks like when an agent offers higher-level cognitive support and picks its own moment to speak.

They built a probe and deployed it with 16 participants for a week. Writers create partners by configuring a role and a proactivity level, and relevant partners then take initiative as the writing happens.

Three findings stand out:

> Participants configured support prospectively, planning for situations they anticipated rather than reacting to interruptions.

> They used suggestions for idea generation and also for self-monitoring, which is a purpose proactive tools rarely design for.

> And they judged intrusiveness by presentation, valuing lightweight visual representation and non-directive rhetorical framing.

The design implications cover customization, timing, engagement and representation.

Worth reading if you are building an assistant that acts before being asked, because how the intervention is phrased mattered to users as much as when it arrived.

Paper: academy.dair.ai/papers/designi…

💬24🔄39❤️257👀21654📊98

"Journey into a Black Hole" One of the many things GPT-6 Astra can do. Just crazy that it can cre...

elvis (@omarsar0) — Sun, 06 Sep 2026 20:09:17 GMT

"Journey into a Black Hole"

One of the many things GPT-6 Astra can do. 

Just crazy that it can create animated scenes like this. 

Uses Vanilla JavaScript, HTML/CSS, and custom WebGL/GLSL shaders.

Your browser does not support the video tag.