Jerry Liu(@jerryjliu0)

LlamaParse can't solve a conjecture, but it can parse this as a PDF 🫡

Jerry Liu (@jerryjliu0) — Thu, 23 Jul 2026 05:24:57 GMT

LlamaParse can't solve a conjecture, but it can parse this as a PDF 🫡

Dmitry Rybin@DmitryRybin1

Dinitz-Garg-Goemans conjecture is false. This graph theory problem was open for ~30 years.

The graph below has fractional flow cost 58. Any unsplittable flow (with capacity violation <=15) has cost at least 60.

Chat with GPT 5.6 Pro where this was found: https://t.co/Oi2PQoab2h

💬5🔄4❤️38👀4352📊10

I have no idea what this is, but I know it’s gonna be good

Jerry Liu (@jerryjliu0) — Thu, 23 Jul 2026 01:46:19 GMT

I have no idea what this is, but I know it’s gonna be good

travis kalanick@travisk

x.com/i/article/2079…

💬2🔄2❤️8👀2533📊3

We're parsing some of the hardest financial documents into clean, plaintext/structured outputs throu...

Jerry Liu (@jerryjliu0) — Wed, 22 Jul 2026 20:04:31 GMT

We're parsing some of the hardest financial documents into clean, plaintext/structured outputs through a live webinar.

Come check it out! watch.getcontrast.io/register/llama…

LlamaIndex 🦙@llama_index

Your agentic workflows are only as good as the context you feed them.

In finance, that context is stuck inside your messiest documents: dense tables, footnoted adjustments, and the details buried in the notes that actually carry the risk. Most tooling flattens exactly the structure that matters, so what reaches your systems is already degraded.

On July 30, join Harsha, a Solution Architect on our team, as he shows how to turn a real, dense financial document into clean, structured, decision-ready context.

On the agenda:
→ Tables and footnotes that stay intact, not flattened into text
→ Extracted figures traceable to the exact source page
→ Values checked against your rules, with exceptions surfaced

Thursday, July 30 · 9:00 AM PT / 12:00 PM ET 
Save your seat: watch.getcontrast.io/register/llama…

💬2🔄2❤️8👀2524📊4

One of the hardest parts of document parsing is getting granular attribution and bounding boxes. Thi...

Jerry Liu (@jerryjliu0) — Wed, 22 Jul 2026 17:31:09 GMT

One of the hardest parts of document parsing is getting granular attribution and bounding boxes. This lets you ground each piece of text in the specific place in the document it came from.

We’ve done some awesome work these past few months improving our bounding boxes, and we now offer multi-layered bounding boxes: region-level, line-level, and word-level. This lets you get attribution at whatever you need - whether the end-user needs to audit a specific number in the invoice, or they want to understand the chart in the research report.

Now available in LlamaParse: cloud.llamaindex.ai

💬4🔄7❤️34👀3534📊13

i used to do this a lot. but then i felt like i was getting progressively dumber at writing. i thi...

Jerry Liu (@jerryjliu0) — Wed, 22 Jul 2026 04:05:25 GMT

i used to do this a lot. but then i felt like i was getting progressively dumber at writing. 

i think there's some value in forcing your own brain to sharpen your thinking, instead of offloading it to AI 

what i found helpful is to manually type out my own understanding in distilled bullet points after reading the LLM output

💬25🔄20❤️331👀52004📊89

It does feel like every frontier lab is optimizing for general intelligence. This leaves room for e...

Jerry Liu (@jerryjliu0) — Wed, 22 Jul 2026 03:15:15 GMT

It does feel like every frontier lab is optimizing for general intelligence.

This leaves room for everyone else to optimize for task-specific intelligence. Every task requires a different point on the cost-capability frontier. Optimizing for each task involves tuning the right set of models and harness for that task to push performance and reduce cost. 

I think there's plenty of room for open-weight models, optimization infrastructure (e.g. RL), vertical-specific workflows, and domain-specific models (e.g. us for document parsing) to thrive in this exponentially growing AI economy

💬29🔄15❤️111👀7950📊53

We dropped some new LlamaDrip 🧢 Fear of Docs LlamaParse

Jerry Liu (@jerryjliu0) — Wed, 22 Jul 2026 01:39:45 GMT

We dropped some new LlamaDrip 🧢

Fear of Docs
LlamaParse

LlamaIndex 🦙@llama_index

The team flew into SF for a week onsite. 🌉

2x'd in size since we last did this — first time this many of us have been in the same room.

The recap:
🧠 Locked in our roadmap for AI agent document infra
🎨 Spraypainting, Topgolf, soccer, & escape rooms
⚽️ Caught a few World Cup matches between sessions
🍗 Raising Cane's for lunch
🎯 Darts + way too many dinners across the city

Ready for Q3.

💬1🔄0❤️16👀4406📊3

We benchmarked Gemini 3.6 Flash and Gemini 3.5 Flash Lite on document understanding. We compared ag...

Jerry Liu (@jerryjliu0) — Wed, 22 Jul 2026 01:33:49 GMT

We benchmarked Gemini 3.6 Flash and Gemini 3.5 Flash Lite on document understanding.

We compared against their prior versions - Gemini 3.5 Flash and Gemini 3.1 Flash Lite.

1️⃣ Gemini 3.6 Flash has roughly similar performance, though it does go down 14% in chart understanding
2️⃣ Gemini 3.5 Flash Lite improves on layout detection by 11%, though it regresses on tables by ~12%

In general it seems like while Gemini 3 Flash was well tuned for document understanding, subsequent versions have been posttrained for coding and reasoning, which has led to a plateau in visual recognition capabilities. 

It would be interesting to see if Google starts to prioritize visual understanding again with the Flash series, or if they're also going all in on reasoning models.

Check out these results and more on ParseBench: parsebench.ai

💬5🔄12❤️131👀14694📊33

the jacobian conjecture post has 3x the impressions of ishowspeed with BTS during the world cup fina...

Jerry Liu (@jerryjliu0) — Tue, 21 Jul 2026 03:42:39 GMT

the jacobian conjecture post has 3x the impressions of ishowspeed with BTS during the world cup final

Speed⭐️@ishowspeedsui

yes i’m a fan🥹🥹🥹🥹

💬6🔄0❤️13👀3892📊6

This is awesome. Instead of doing work this morning, I spent some time brushing up on math to provid...

Jerry Liu (@jerryjliu0) — Mon, 20 Jul 2026 17:46:44 GMT

This is awesome. Instead of doing work this morning, I spent some time brushing up on math to provide a handwavy overview to anyone else who like me doesn't have a math background and is rusty: 

At a distance, the conjecture is basically that if a n-dimensional polynomial function (e.g. F(x, y, z) map to some output (x', y', z')) meets some special conditions, then 1) it is invertible, and 2) the inverse is polynomial. 

Disproving the conjecture requires finding a n-dimensional polynomial function that meets these special conditions, but is either 1) not invertible at all, or 2) is invertible but not polynomial. (2) is actually not a possibility due to some other math theorem - if there is an inverse, the inverse has to be polynomial. 

Invertibility just means the function can be "undone" at every single input point - if input (x1, y1, z1) maps to output (x2, y2, z2), then output (x2, y2, z2) can map to input (x1, y1, z1). A function is not invertible if multiple input points (x, y, z)'s map to the same output (x, y, z)

Here Fable 5 presumably did some search and found a function that was not invertible. 

That function is this:  
F(x, y, z) = (
z(1+xy)^3 + y^2(1+xy)(4+3xy),
y+3x(1+xy)^2z + 3xy^2(4+3xy),
2x-3x^2y-x^3z
)

This function meets the special conditions (jacobian determinant is a non-zero constant for every point), but has multiple input points that map to the same output: 

Inputs: (0, 0, -1/4), (1, -3/2, 13/2), and (-1, 3/2, 13/2)
all map to the same output: (-1/4, 0, 0)

-----

The Jacobian is the multivariable version of a derivative, so it is a function of a specific input point. When the Jacobian determinant is non-zero at a specific input point, there is local invertibility around that point. This conjecture was stronger in that if it was non-zero (and constant, due to being over the complex numbers, which I don't quite get and I ran out of time to understand), then the function is globally invertible. 

-----

What's interesting here is that finding counterexamples seem like the type of math work that agentic models excel at. It involves generating and crawling a large search space, using priors to narrow it down, and a bit of brute force and raw compute power. Machines can process a wide set of possibilities faster than humans, which is helpful for finding counterexamples.

levent@__alpoge__

hello there the jacobian conjecture is false thanx to my close friend akhil for asking about it and my other close friend fable for working during the world cup final

((1+xy)^3 z + y^2 (1+xy) (4+3xy), y + 3 x (1+xy)^2  z + 3 x y^2 (4+3xy), 2 x - 3 x^2 y - x^3 z): \C^3\to \C^3, has jacobian determinant -2, and sends (0, 0, -1/4), (1, -3/2, 13/2), and (-1, 3/2, 13/2) to (-1/4, 0, 0)

💬12🔄6❤️140👀20145📊33

Hosting a small founder dinner in sf on july 28 with @dexhorthy - first in a monthly series. Topic ...

Jerry Liu (@jerryjliu0) — Mon, 20 Jul 2026 17:00:22 GMT

Hosting a small founder dinner in sf on july 28 with @dexhorthy - first in a monthly series.

Topic is loop engineering: context management, compaction, evals, human-in-the-loop handoffs.

Approval-only and most seats are spoken for, but if you're a founder building agents in production: luma.com/llamai-c6of

💬7🔄2❤️21👀2585📊10

I don't think there's a ton of value in creating custom general harnesses. But there will always tr...

Jerry Liu (@jerryjliu0) — Sun, 19 Jul 2026 17:23:42 GMT

I don't think there's a ton of value in creating custom general harnesses.

But there will always tremendous value in creating custom task-specific harnesses, especially if you need to satisfy accuracy/cost/latency constraints for production. You can encode all the domain-specific priors, the right model mixture, and the right UX so it's far better than you trying to prompt a general agent with some skills/prompts. 

1. You can always bootstrap any workflow with Codex/Claude Code first
2. Distill it into something more specific

Ishan@radshaan

Is there even a point in creating a custom harness anymore? I feel like Codex/Claude Code/Open Code/etc. have all implemented the hard parts correctly and we should just use coding harnesses as general purpose harnesses

💬21🔄4❤️101👀12404📊37

We've built the following document retrieval endpoints into LlamaParse: * Hybrid search (grep + vec...

Jerry Liu (@jerryjliu0) — Sun, 19 Jul 2026 16:59:53 GMT

We've built the following document retrieval endpoints into LlamaParse: 
* Hybrid search (grep + vector search) 
* File grep (grep incl. regex search) 
* File find (`find`) 
* File read (`sed`) 

We're continuing to experiment to see what combination allows agents to get the highest retrieval quality over unstructured docs. If you have thoughts let us know!

Jerry Liu@jerryjliu0

I'm glad people still understand the importance of building high-quality retrieval systems in 2026, especially as the outer models/harnesses are getting better every day.

Making agentic retrieval work in production  doesn't necessarily require groundbreaking new techniques around retrieval or planning. This article shows that you do need to spend engineering time tuning chunking, synchronization, reranking, tool API design, permissioning, and more. Some interesting tidbits:
* Concatenating onto existing Slack threads as one contiguous chunk with heuristics for determining additional relevant context
* Real-time updates for Slack
* Separate chunking/updates for codebases
* The discovery that hybrid search works well (not like this was a super new realization, but always good to validate and understand the specific parameters they tweaked) 
* Having natural guardrails on which data sources are relevant for which projects

Building simple retrieval is easy, building production retrieval is hard. We had to deal with a lot of these challenges when productionizing our own Index feature within LlamaParse.

💬19🔄7❤️57👀6302📊28

I've noticed this too. I'm starting to have trouble understanding and articulating what AI generates

Jerry Liu (@jerryjliu0) — Sun, 19 Jul 2026 16:37:44 GMT

I've noticed this too. I'm starting to have trouble understanding and articulating what AI generates

Ashwinee Panda@PandaAshwinee

the fable / sol summaries are getting more and more absurdly information-dense
1 in every 5 words is a novel compound word (ex: “rate-adjacent”) which itself is the model’s invented stand-in for a metaconcept that it developed that i need to interrogate it about
how to fix this?

💬30🔄6❤️182👀44672📊49

Palantir might be one of the best positioned companies to sell RL envs to frontier labs

Jerry Liu (@jerryjliu0) — Sat, 18 Jul 2026 18:09:56 GMT

Palantir might be one of the best positioned companies to sell RL envs to frontier labs

💬32🔄6❤️323👀92929📊85

We've invested a lot of engineering resources in the past 2 years making our own indexing pipelines ...

Jerry Liu (@jerryjliu0) — Sat, 18 Jul 2026 17:48:50 GMT

We've invested a lot of engineering resources in the past 2 years making our own indexing pipelines reliable, accurate, and agent-native with a variety of retrieval endpoints designed to plug in to various harnesses.

Check it out: cloud.llamaindex.ai

💬5🔄1❤️7👀1894📊7

I'm glad people still understand the importance of building high-quality retrieval systems in 2026, ...

Jerry Liu (@jerryjliu0) — Sat, 18 Jul 2026 17:48:49 GMT

I'm glad people still understand the importance of building high-quality retrieval systems in 2026, especially as the outer models/harnesses are getting better every day.

Making agentic retrieval work in production  doesn't necessarily require groundbreaking new techniques around retrieval or planning. This article shows that you do need to spend engineering time tuning chunking, synchronization, reranking, tool API design, permissioning, and more. Some interesting tidbits:
* Concatenating onto existing Slack threads as one contiguous chunk with heuristics for determining additional relevant context
* Real-time updates for Slack
* Separate chunking/updates for codebases
* The discovery that hybrid search works well (not like this was a super new realization, but always good to validate and understand the specific parameters they tweaked) 
* Having natural guardrails on which data sources are relevant for which projects

Building simple retrieval is easy, building production retrieval is hard. We had to deal with a lot of these challenges when productionizing our own Index feature within LlamaParse.

Cerebras@cerebras

x.com/i/article/2077…

💬16🔄30❤️265👀49421📊94

1. build a workflow graph on top of the agent harness 2. dynamically create that graph through an ou...

Jerry Liu (@jerryjliu0) — Sat, 18 Jul 2026 16:59:07 GMT

1. build a workflow graph on top of the agent harness
2. dynamically create that graph through an outer agent loop
....repeat

life is a loop (and a graph)

Peter Steinberger 🦞@steipete

Are we still talking loops or did we shift to graphs yet?

💬20🔄8❤️185👀25765📊61

Tsinghua is the MIT of china, xi jingping went there, everyone i've met from there is cracked (inclu...

Jerry Liu (@jerryjliu0) — Sat, 18 Jul 2026 16:10:41 GMT

Tsinghua is the MIT of china, xi jingping went there, everyone i've met from there is cracked (including my parents lol)

💬39🔄32❤️792👀158652📊159

you know what, i'll take it (as a max user 🙏)

Jerry Liu (@jerryjliu0) — Sat, 18 Jul 2026 02:58:43 GMT

you know what, i'll take it 

(as a max user 🙏)

Claude@claudeai

Beginning July 20, Claude Fable 5 will be included in all Max and Team Premium plans, at 50% of limits.

Pro and Team Standard users will continue to have access to Fable via usage credits, and will receive a one-time $100 credit.

Demand for Fable has been challenging to predict, which is why we rolled it out to subscription plans in stages, extending access several times as we secured additional capacity.

💬3🔄0❤️32👀8119📊5

this is just twitter

Jerry Liu (@jerryjliu0) — Sat, 18 Jul 2026 00:24:19 GMT

this is just twitter

brett goldstein@thatguybg

we need BookFace for non-YC startups

founder friend got ghosted for 6 weeks
after 4 meetings and a bunch of customer diligence convos

VC only responded when a portco founder nudged them, but suggested she meet with an analyst (and answer same questions again)

she said no obvi

💬5🔄0❤️14👀5695📊6

We have a new company-wide swag collection 🔥

Jerry Liu (@jerryjliu0) — Fri, 17 Jul 2026 22:29:58 GMT

We have a new company-wide swag collection 🔥

Jerry Liu@jerryjliu0

This week we had a huge global company onsite in SF 🌉 (the company has 2x’ed in size since our last one!) We solidified our roadmap towards becoming the best document infrastructure for AI agents. We’re building a cracked AI lab focused on all aspects of document work - from document OCR, extraction, search, to e2e workflows - and hillclimbing our capabilities through model training, agent engineering, and deep infrastructure expertise.

We’re also committed to building the most comprehensive, unbiased benchmarks to advance the Pareto frontier for all document tasks. ParseBench has 2k+ enterprise pages, ~80 solutions verified, 15k monthly downloads, and is available across HuggingFace and Kaggle. Next up is increasing our dataset (more verticals) and task coverage (e.g extraction). LlamaParse agentic and cost-effective modes directly define the Pareto curve for VLM-based OCR solutions and we’re making rapid leaps across complex tables, charts, fonts, layouts and more every single day. In addition, liteparse is the fastest, most performant free parser out there and is quickly becoming one of the most popular, eclipsing 11.5k+ stars in ~2 months.

This week we did some work and had a lot of fun. Maybe a little too much fun 😅:
* Spraypainting
* Darts
* Topgolf
* Escape Room
* Soccer
* Many dinners
* Raising Canes for lunch

Time to lock back in. You’re going to see us ship an insane amount of stuff in Q3.

Your browser does not support the video tag.

💬3🔄0❤️18👀4979📊4

Inspired from some of the hoodies I personally own. Thanks @murtazakhomusi for humoring me on this

Jerry Liu (@jerryjliu0) — Fri, 17 Jul 2026 22:29:58 GMT

Inspired from some of the hoodies I personally own. Thanks @murtazakhomusi for humoring me on this

💬1🔄0❤️5👀813📊2

This week we had a huge global company onsite in SF 🌉 (the company has 2x’ed in size since our last...

Jerry Liu (@jerryjliu0) — Fri, 17 Jul 2026 21:56:04 GMT

This week we had a huge global company onsite in SF 🌉 (the company has 2x’ed in size since our last one!) We solidified our roadmap towards becoming the best document infrastructure for AI agents. We’re building a cracked AI lab focused on all aspects of document work - from document OCR, extraction, search, to e2e workflows - and hillclimbing our capabilities through model training, agent engineering, and deep infrastructure expertise.

We’re also committed to building the most comprehensive, unbiased benchmarks to advance the Pareto frontier for all document tasks. ParseBench has 2k+ enterprise pages, ~80 solutions verified, 15k monthly downloads, and is available across HuggingFace and Kaggle. Next up is increasing our dataset (more verticals) and task coverage (e.g extraction). LlamaParse agentic and cost-effective modes directly define the Pareto curve for VLM-based OCR solutions and we’re making rapid leaps across complex tables, charts, fonts, layouts and more every single day. In addition, liteparse is the fastest, most performant free parser out there and is quickly becoming one of the most popular, eclipsing 11.5k+ stars in ~2 months.

This week we did some work and had a lot of fun. Maybe a little too much fun 😅:
* Spraypainting
* Darts
* Topgolf
* Escape Room
* Soccer
* Many dinners
* Raising Canes for lunch

Time to lock back in. You’re going to see us ship an insane amount of stuff in Q3.

Your browser does not support the video tag.

💬4🔄2❤️21👀7910📊8

This is absolutely insane. Having a huge proliferation of Chinese open-source models that are winnin...

Jerry Liu (@jerryjliu0) — Fri, 17 Jul 2026 15:35:55 GMT

This is absolutely insane. Having a huge proliferation of Chinese open-source models that are winning on task-specific benchmarks may not be great for American AI dominance. But - it is fantastic news for every non-frontier lab out there.

Every American AI startup will survive by distilling from the latest open-weight models, instead of solely depending on the frontier labs.

Arena.ai@arena

Big news: Kimi-K3 by @Kimi_Moonshot is now #1 in the Frontend Code Arena with 1679 pts, surpassing Claude Fable 5.

This is a 17-place jump from Kimi-k2.6 (#18 -> #1).

In Frontend, Kimi-K3 ranked #1 in 6 of 7 domains: Brand & Marketing, Reference-Based Design, Data & Analytics, Consumer Product, Simulations, and Content Creation Tools, landing #2 only in Gaming behind Fable 5.

The full model weights will be released by July 27.

Congrats to the @Kimi_Moonshot team on this major milestone!

💬13🔄4❤️39👀5282📊16

We Parse PDFs 📑 company photoshoot

Jerry Liu (@jerryjliu0) — Thu, 16 Jul 2026 23:27:13 GMT

We Parse PDFs 📑 company photoshoot

Murtaza Khomusi@murtazakhomusi

vibez are g00d @llama_index

Your browser does not support the video tag.

💬3🔄1❤️9👀3008📊4

I do think that for any given task, you can always distill a generalized model/agent harness into a ...

Jerry Liu (@jerryjliu0) — Thu, 16 Jul 2026 17:18:16 GMT

I do think that for any given task, you can always distill a generalized model/agent harness into a specialized model/harness/workflow for higher accuracy and lower cost.

The more narrow the data and task domain, the more opportunity there is to optimize performance (not to mention provide a more targeted UX vs. a chat interface). 

As agents get better at improving themselves, I think you can start to perform this distillation automatically, by defining a top-level goal and evaluation rubric, instead of codifying the workflow through code or natural language.

Jerry Liu@jerryjliu0

x.com/i/article/2077…

💬12🔄13❤️106👀11318📊42

Parse PDFs at lightspeed through grpc liteparse already had a REST API, but now we've built in sup...

Jerry Liu (@jerryjliu0) — Thu, 16 Jul 2026 17:06:47 GMT

Parse PDFs at lightspeed through grpc 

liteparse already had a REST API, but now we've built in support for liteparse as a grpc service with typed fields through protobufs. This is great if you're looking to have fast, efficient document parsing within your backend services. 

Check it out: llamaindex.ai/blog/introduci…

LlamaIndex 🦙@llama_index

Document processing pipelines are often made up of multiple services: parsing → chunking → embeddings → storage.

Until now, LiteParse was only available over REST. That works well for browsers and simple integrations, but it’s not always the best fit for service-to-service communication.

Today we’re introducing liteparse-grpc.

With  @𝘭𝘭𝘢𝘮𝘢𝘪𝘯𝘥𝘦𝘹/𝘭𝘪𝘵𝘦𝘱𝘢𝘳𝘴𝘦-𝘨𝘳𝘱𝘤 (npm) or the  𝘨𝘩𝘤𝘳.𝘪𝘰/𝘳𝘶𝘯-𝘭𝘭𝘢𝘮𝘢/𝘭𝘪𝘵𝘦𝘱𝘢𝘳𝘴𝘦-𝘨𝘳𝘱𝘤 Docker image, you can run LiteParse as a gRPC service and:

📄 Parse PDFs, Office documents, and images into JSON, text, or Markdown
📸 Render PDF pages as screenshots
📊 Estimate document complexity and whether OCR is needed

The package also includes a CLI client for testing, generated TypeScript client stubs and the protobuf definitions, so you can generate clients for Python, Go, Java, or any other language supported by 𝘱𝘳𝘰𝘵𝘰𝘤

The REST server isn’t going anywhere, this is simply another interface for teams building polyglot, microserllamaindex.ai/blog/introduci…elines.

📖github.com/run-llama/lite…n: https://t.co/iAIBoXzhEH
💻 GitHub: https://t.co/K0d8XVE7Rc

💬6🔄6❤️36👀5101📊15

Jerry Liu (@jerryjliu0) — Thu, 16 Jul 2026 04:24:14 GMT

ClaudeDevs@ClaudeDevs

We've reset 5-hour and weekly rate limits for all users.

💬5🔄0❤️26👀3935📊6

https://t.co/FLiVbGVK1a

Jerry Liu (@jerryjliu0) — Wed, 15 Jul 2026 23:36:36 GMT

x.com/i/article/2077…

💬14🔄21❤️243👀44875📊134

Extracting information from millions of documents at scale used to take an insane number of human ho...

Jerry Liu (@jerryjliu0) — Wed, 15 Jul 2026 16:57:53 GMT

Extracting information from millions of documents at scale used to take an insane number of human hours. Even with recent OCR + document AI tech, humans would still have to spend a lot of time carefully tuning the schema and precisely defining the fields they want to extract.

We've launched a Conversational Extract feature within LlamaParse that now lets you define and run complex document extraction through a conversation. 

1️⃣ Describe the fields you want to extract, and upload a  reference document
2️⃣ Refine the document schema directly, or through additional conversations. Select the right level of agentic reasoning needed
3️⃣ Run extraction at scale over 1M+ docs. Get bounding boxes, citations, and confidence scores for every single extracted field.

Check it out today: cloud.llamaindex.ai

Your browser does not support the video tag.

LlamaIndex 🦙@llama_index

Your extraction schema is now a conversation away.
⁣
📄 Upload a doc — the agent drafts the schema from it⁣
💬 Want changes? Just ask⁣
📚 Or grab a template and go⁣
⁣
Writing JSON Schema by hand? That's over.⁣
⁣
Conversational Extract is live in LlamaParse. Try it now cloud.llamaindex.ai/?utm_medium=so…ip6

Your browser does not support the video tag.

💬7🔄7❤️80👀11704📊26

enemy of my enemy is my friend type shit

Jerry Liu (@jerryjliu0) — Wed, 15 Jul 2026 03:27:08 GMT

enemy of my enemy is my friend type shit

Sam Altman@sama

this is a thoughtful proposal from demis:

💬8🔄2❤️32👀12287📊10

Work hard play hard 🇫🇷🇪🇸

Jerry Liu (@jerryjliu0) — Tue, 14 Jul 2026 20:03:02 GMT

Work hard play hard 🇫🇷🇪🇸

Your browser does not support the video tag.

💬3🔄0❤️31👀6560📊6

100%. If I'm driving on the 101 it is a life hazard to read verbose billboards

Jerry Liu (@jerryjliu0) — Tue, 14 Jul 2026 19:55:23 GMT

100%. If I'm driving on the 101 it is a life hazard to read verbose billboards

Murtaza Khomusi@murtazakhomusi

SF peepz. Stop trying to fit more than 5 words on a billboard. Bad vibez. 

Thank you for your attention to this matter.

💬4🔄0❤️9👀4709📊4

If you're a founder, come pitch your wildest ideas to other founders in a safe space! Food and dri...

Jerry Liu (@jerryjliu0) — Tue, 14 Jul 2026 17:09:03 GMT

If you're a founder, come pitch your wildest ideas to other founders in a safe space! 

Food and drinks included 🍹

LlamaIndex 🦙@llama_index

Co-hosting 'Swap and Pitch Night' this Thursday in SF with @tembo, @clerk, and @autonomaai. 

Pitch night where founders pitch each other's startups. Good test of whether you actually listen, and sometimes someone pitches your startup better than you do.

855 Brannan, 6 to 8:30pm. 

RSVP here: luma.com/swapandpitch

💬4🔄1❤️15👀4419📊5

Our team made the cutest graffiti art yesterday 🎨

Jerry Liu (@jerryjliu0) — Tue, 14 Jul 2026 17:03:03 GMT

Our team made the cutest graffiti art yesterday 🎨

LlamaIndex 🦙@llama_index

We eat docs🫢

💬6🔄1❤️19👀3176📊7

Agreed with this. people underestimate the importance of good abstractions and maintainability. pa...

Jerry Liu (@jerryjliu0) — Tue, 14 Jul 2026 00:55:03 GMT

Agreed with this. people underestimate the importance of good abstractions and maintainability. 

part of the reason LLMs/AI are so popular in the first place is because of the ease of using the models during inference-time (through prompting, orchestration) instead of training

Jesse Zhang@thejessezhang

I actually feel strongly that the "learning" companies will want to do with their data will mostly not be to train models.

Let's say you have a bunch of valuable data about your customers or employees.

Your best bet to make that IP useful is to turn it into skills or artifacts that models can use in-context.

If you go through the effort to train it into the model:

1. that is hard to get right and takes a lot of time + effort
2. you will have to do it all over again when the next base model comes out
3. most importantly, it is irreversible... when things change, you cannot untrain what you did

💬15🔄3❤️31👀6127📊18

OpenAI is absolutely cooking this week

Jerry Liu (@jerryjliu0) — Mon, 13 Jul 2026 22:21:33 GMT

OpenAI is absolutely cooking this week

Sam Altman@sama

come for the best model, stay because we don’t treat you with contempt

💬6🔄1❤️38👀6903📊9

There is a massive opportunity for any startup to build specialized, domain-specific workflows that ...

Jerry Liu (@jerryjliu0) — Mon, 13 Jul 2026 22:10:10 GMT

There is a massive opportunity for any startup to build specialized, domain-specific workflows that are at the Pareto frontier in accuracy, cost, and latency.

It is clear that not every task requires frontier intelligence (e.g. 5.6 Sol xhigh). Every single task has an optimal model+harness configuration in order to solve that task in an accurate, cheap, fast manner. This holds true even as frontier models get better!

We see this as being especially true for document OCR. We think there will always be a massive gap in the Pareto frontier for document OCR vs. frontier model capabilities for visual understanding.

Over the past year, we’ve invested massive amounts of research efforts in building auto-routing harnesses, fine-tuned layout models, and specialized models for specific elements within the document distribution - tables, charts, handwriting, and more.

All this applied research makes its way towards making LlamaParse the best document processing service out there. Come check it out: cloud.llamaindex.ai

💬14🔄21❤️182👀13217📊77

Parse any PDF into agent-ready context. Fully local, free desktop app. No model / internet connectio...

Jerry Liu (@jerryjliu0) — Mon, 13 Jul 2026 20:35:48 GMT

Parse any PDF into agent-ready context. Fully local, free desktop app. No model / internet connection required!

LlamaIndex 🦙@llama_index

Ever wanted to quickly turn a PDF into clean text to paste into your favorite AI agent, without having use CLIs or open the browser?

We built exactly that.

Using @TauriAp ps, with a Rust backend powered by LiteParse and a React frontend, we created a cross-platform desktop app that lets you:

• Drag and drop PDFs
• Convert them into clean Markdown
• Preview page screenshots with extracted bounding boxes overlaid for easy inspection

It's a simple way to understand your document's structure and verify the parsing results before handing them off to an LLM.

Check out the demo below! 👇

GitHub:github.com/run-llama/lite…W
Get started with LiteParse:developers.llamaindex.ai/liteparse/S

Your browser does not support the video tag.

💬2🔄5❤️66👀13551📊22

congrats on the launch! As the task horizons on agents gets longer, we need more work evaluating a...

Jerry Liu (@jerryjliu0) — Mon, 13 Jul 2026 18:03:11 GMT

congrats on the launch! 

As the task horizons on agents gets longer, we need more work evaluating and training models to be better in long-running, open-ended, evolving real-world environments.

Skyfall AI@skyfallai

Today we present Morpheus, a persistent enterprise simulation platform designed to make Continual Learning a reality. Morpheus is the world’s first real world Reinforcement Learning environment.

Every Reinforcement Learning environment operates in the game world. Benchmarks like Atari, OpenAI Gym, MuJoCo, and Procgen are all small, game-like worlds that reset every few minutes.

But the real world never resets. A business keeps running and evolving everyday.

We tested how frontier LLMs would perform in realistic and dynamic business environments 🧬on Morpheus. The main conclusion was that LLMs are not continual learners.

🧵Here’s how we did it and what we learned:

Your browser does not support the video tag.

💬9🔄3❤️25👀6244📊13

lmao

Jerry Liu (@jerryjliu0) — Sun, 12 Jul 2026 17:32:49 GMT

lmao

Claude@claudeai

We're extending Claude Fable 5 access on all paid plans, as well as keeping Claude Code’s weekly rate limits 50% higher, through July 19.

💬7🔄0❤️46👀12585📊10

sama scrolling AI Twitter like the rest of us

Jerry Liu (@jerryjliu0) — Sat, 11 Jul 2026 18:17:43 GMT

sama scrolling AI Twitter like the rest of us

Sam Altman@sama

30% of the cost was on fable at these levels of usage?

💬9🔄3❤️98👀22441📊17

wtf I thought this was AI cc @disiok

Jerry Liu (@jerryjliu0) — Fri, 10 Jul 2026 03:05:28 GMT

wtf I thought this was AI 

cc @disiok

Lawrence Jang@JangLawrenceK

guys I just cancelled my Claude plan I don’t know what happened

Your browser does not support the video tag.

💬6🔄1❤️17👀12633📊8

We comprehensively benchmarked GPT-5.6 on document understanding. At a high-level there's no change...

Jerry Liu (@jerryjliu0) — Thu, 09 Jul 2026 23:07:56 GMT

We comprehensively benchmarked GPT-5.6 on document understanding.

At a high-level there's no change between GPT-5.6 Sol and GPT-5.5 in terms of performance over tables, text, charts, layout, and more. The GPT-class of models typically does quite well over table understanding, but they struggle with transcribing complex text layouts and formatting, with transcribing charts, and with deriving bounding boxes over source elements.

Come check out our leaderboard of over 70+ frontier models, open-weight models, and OCR solutions on ParseBench: parsebench.ai

LlamaIndex 🦙@llama_index

@OpenAI  released GPT 5.6 today and we ran a day 0 benchmark in ParseBench to test improvements in document understanding. The new family of models continues to excel at reading text and tables, but continues to struggle with charts and layout. 

What's most interesting is that Luna is about 6 x cheaper than Sol and only results in minor degradations across all ParseBench metrics, showing that an increase in reasoning tokens doesn't always result in a commensurate improvement in visual understanding.

💬12🔄10❤️115👀19233📊27

Actually I did a quick test. I like the ChatGPT Work/Codex split better than Claude Cowork/Code. T...

Jerry Liu (@jerryjliu0) — Thu, 09 Jul 2026 19:12:18 GMT

Actually I did a quick test. I like the ChatGPT Work/Codex split better than Claude Cowork/Code. 

The interface is much more unified. The functionality is effectively the same. The chat history is shared between the two modes. The team put more care into the specific toggles that should feel different between knowledge work and code.

Jerry Liu@jerryjliu0

ChatGPT Work == Claude Cowork 
ChatGPT Codex == Claude Code

I kinda wish OpenAI created a single unified app surface for all work, coding or not, even though I get the UI/UX would be different

💬6🔄2❤️35👀6231📊9

ChatGPT Work == Claude Cowork ChatGPT Codex == Claude Code I kinda wish OpenAI created a single un...

Jerry Liu (@jerryjliu0) — Thu, 09 Jul 2026 19:06:09 GMT

ChatGPT Work == Claude Cowork 
ChatGPT Codex == Claude Code

I kinda wish OpenAI created a single unified app surface for all work, coding or not, even though I get the UI/UX would be different

OpenAI@OpenAI

Introducing ChatGPT Work, a new agent in ChatGPT powered by Codex and GPT-5.6.

It can take action across your apps and files, stay with a project for hours if needed, and turn a goal into finished work.

It’s a whole new way to get work done.

Your browser does not support the video tag.

💬10🔄3❤️21👀9230📊11

Congrats to our llama cousins 🫡🦙

Jerry Liu (@jerryjliu0) — Thu, 09 Jul 2026 15:02:45 GMT

Congrats to our llama cousins 🫡🦙

ollama@ollama

Big day for Ollama! When we started, open models and the open source AI ecosystem were in their early days with few believers.

Our belief in open source has never wavered.

With today's fundraising announcement and our 9M+ active builders, we’re ready to scale open models into AI that you can own.

All aboard open models!

🧵

💬2🔄0❤️13👀3295📊3