Jerry Liu(@jerryjliu0)

This week we launched the world's most accurate document extraction agent over real-world documents....

Jerry Liu (@jerryjliu0) — Thu, 13 Aug 2026 13:46:38 GMT

This week we launched the world's most accurate document extraction agent over real-world documents.

Introducing LlamaExtract Agentic Plus 💫 . It is a complete document extraction model+harness engine that can convert even the most complex document types/lengths/schemas into clean, grounded, structured output.

We benchmarked it over ExtractBench - a comprehensive set of real-world enterprise documents. It is measured over various dimensions: document length (short/medium/long), task (long-list, needle in haystack, dense docs), perception (rotated, scanned, handwriting), table structured (merged headers, cross-page tables, table within cells), and business domains.

We compared it against a variety of coding agent harness to document extraction solutions. LlamaExtract Agentic Plus is not only highly accurate, it is quite price efficient relative to its capabilities (25-50% the price of Claude Code/Codex).

You can see the full release of ExtractBench here:llamaindex.ai/blog/introduci…z

If you want to check it out, come sign up for an account on LlamaParse:cloud.llamaindex.aiA

Jerry Liu@jerryjliu0

Introducing ExtractBench, the most comprehensive benchmark for information extraction from complex enterprise documents.

The latest models are pushing the frontier of coding and knowledge work, but surprisingly they still struggle on complex doc extraction tasks in production. A well-tuned extractor must parse multi-page filings without dropping rows, emit exact spatial citations for auditability, and handle messy scans. Also they must do all of this at a viable per-page cost so that you can scale this to millions of docs in production (you can’t be paying upwards of $1 in tokens per page!)

Existing extraction benchmarks fall short: they are not large/diverse enough in document domain (finance, energy, gov, auto), elements (long records, scans, grounding), and schemas.

So our applied research team built ExtractBench.

We evaluated 14 systems: frontier VLMs, coding agents, and specialized extraction APIs, against 370 enterprise documents: 4,869 pages, 67 document types.

Our biggest finding 🧪: Short documents mask critical system flaws. On files past 50 pages, commercial VLMs collapse below 35% recall due to silent list truncation. They hold high precision, but lose output attention and drop most of the table rows.

ExtractBench evaluates value accuracy, long-record completeness, spatial grounding, and per-page cost with zero LLM judges. It is 100% deterministic and reproducible.

In tandem with ExtractBench, we’re also introducing 𝗔𝗴𝗲𝗻𝘁𝗶𝗰 𝗣𝗹𝘂𝘀, a new Extract tier in LlamaParse that debuts at #1 on the leaderboard: 95.6% value accuracy, at less than a third the cost of the closest peer.

Explore the findings, download the dataset, or run the harnellamaindex.ai/blog/introduci…o/tsXfovKgithub.com/run-llama/Extr…o/auzmOGgUl3
Hhuggingface.co/datasets/llama…o/1eKisSBrE8

We will be actively evolving both our extraction benchmark as well as our extraction harness over time. 

If you check out either ExtractBench or LlamaParse, let us know your feedback!

Your browser does not support the video tag.

💬7🔄6❤️58👀6673📊17

If you’re interested in checking out LlamaParse for document extraction, sign up here: https://t.co/...

Jerry Liu (@jerryjliu0) — Wed, 12 Aug 2026 15:18:59 GMT

If you’re interested in checking out LlamaParse for document extraction, sign up here: login.llamaindex.ai/sign-up

💬1🔄0❤️1👀1026📊2

We wrote a 36-page ArXiv whitepaper on ExtractBench 🧑‍🔬 , our effort to create the most comprehens...

Jerry Liu (@jerryjliu0) — Wed, 12 Aug 2026 15:18:58 GMT

We wrote a 36-page ArXiv whitepaper on ExtractBench 🧑‍🔬 , our effort to create the most comprehensive, schema-guided, real-world document extraction benchmark.

It’s extremely detailed and covers everything from comparisons with related work on document extraction, to the dataset construction / how ground-truth is generated, to our experiments over 14+ extraction systems.

Here are some of the most salient points from the paper:
✅ The benchmark scores schema-guided extraction on real enterprise documents. Given input doc + schema and predicted output from an extractor, the benchmark measures value accuracy, grounding, tags for each “challenge”, and cost.
✅ Compared to other benchmarks, we have more schemas, more evaluation dimensions, and more data domain diversity
✅ The ground-truth is constructed according to 3 doc subtypes: real-docs use a model ensemble + human review, synthetic long lists have ground-truth by construction, and scanned forms also have human review incl. boxes
✅ Our three modes (LlamaParse cost-effective, agentic, and agentic plus) are at the Pareto frontier of accuracy and cost.

The full ArXiv paper is herearxiv.org/pdf/2607.29677zr

Our siteextractbench.aic3

Jerry Liu@jerryjliu0

Introducing ExtractBench, the most comprehensive benchmark for information extraction from complex enterprise documents.

The latest models are pushing the frontier of coding and knowledge work, but surprisingly they still struggle on complex doc extraction tasks in production. A well-tuned extractor must parse multi-page filings without dropping rows, emit exact spatial citations for auditability, and handle messy scans. Also they must do all of this at a viable per-page cost so that you can scale this to millions of docs in production (you can’t be paying upwards of $1 in tokens per page!)

Existing extraction benchmarks fall short: they are not large/diverse enough in document domain (finance, energy, gov, auto), elements (long records, scans, grounding), and schemas.

So our applied research team built ExtractBench.

We evaluated 14 systems: frontier VLMs, coding agents, and specialized extraction APIs, against 370 enterprise documents: 4,869 pages, 67 document types.

Our biggest finding 🧪: Short documents mask critical system flaws. On files past 50 pages, commercial VLMs collapse below 35% recall due to silent list truncation. They hold high precision, but lose output attention and drop most of the table rows.

ExtractBench evaluates value accuracy, long-record completeness, spatial grounding, and per-page cost with zero LLM judges. It is 100% deterministic and reproducible.

In tandem with ExtractBench, we’re also introducing 𝗔𝗴𝗲𝗻𝘁𝗶𝗰 𝗣𝗹𝘂𝘀, a new Extract tier in LlamaParse that debuts at #1 on the leaderboard: 95.6% value accuracy, at less than a third the cost of the closest peer.

Explore the findings, download the dataset, or run the harnellamaindex.ai/blog/introduci…o/tsXfovKgithub.com/run-llama/Extr…o/auzmOGgUl3
Hhuggingface.co/datasets/llama…o/1eKisSBrE8

We will be actively evolving both our extraction benchmark as well as our extraction harness over time. 

If you check out either ExtractBench or LlamaParse, let us know your feedback!

Your browser does not support the video tag.

💬11🔄13❤️134👀15421📊42

ExtractBench is one of the most comprehensive benchmarks for real-world document extraction. ✅ It ...

Jerry Liu (@jerryjliu0) — Tue, 11 Aug 2026 22:16:28 GMT

ExtractBench is one of the most comprehensive benchmarks for real-world document extraction. 

✅ It covers 4869 pages, across 67 document types, spanning 8 real-world domains: finance, energy, gov, auto, supply chain, healthcare, legal, real estate
✅ It covers a distribution of short, medium, and long documents
✅ It covers a variety of very complex table edge cases: tables with over 1k rows, nested tables within cells, cross-page tables, and more
✅ It covers scans, handwriting, and rotated pages

We benchmarked across 14 different VLMs, coding agents, and document extraction APIs.

Everything is fully public on our blog, ArXiv, Github, and HuggingFace: 

Blog: llamaindex.ai/blog/introduci…
Hugging Face: huggingface.co/datasets/llama… 
Github: github.com/run-llama/Extr…
ArXiv: arxiv.org/pdf/2607.29677

Jerry Liu@jerryjliu0

Introducing ExtractBench, the most comprehensive benchmark for information extraction from complex enterprise documents.

The latest models are pushing the frontier of coding and knowledge work, but surprisingly they still struggle on complex doc extraction tasks in production. A well-tuned extractor must parse multi-page filings without dropping rows, emit exact spatial citations for auditability, and handle messy scans. Also they must do all of this at a viable per-page cost so that you can scale this to millions of docs in production (you can’t be paying upwards of $1 in tokens per page!)

Existing extraction benchmarks fall short: they are not large/diverse enough in document domain (finance, energy, gov, auto), elements (long records, scans, grounding), and schemas.

So our applied research team built ExtractBench.

We evaluated 14 systems: frontier VLMs, coding agents, and specialized extraction APIs, against 370 enterprise documents: 4,869 pages, 67 document types.

Our biggest finding 🧪: Short documents mask critical system flaws. On files past 50 pages, commercial VLMs collapse below 35% recall due to silent list truncation. They hold high precision, but lose output attention and drop most of the table rows.

ExtractBench evaluates value accuracy, long-record completeness, spatial grounding, and per-page cost with zero LLM judges. It is 100% deterministic and reproducible.

In tandem with ExtractBench, we’re also introducing 𝗔𝗴𝗲𝗻𝘁𝗶𝗰 𝗣𝗹𝘂𝘀, a new Extract tier in LlamaParse that debuts at #1 on the leaderboard: 95.6% value accuracy, at less than a third the cost of the closest peer.

Explore the findings, download the dataset, or run the harnellamaindex.ai/blog/introduci…o/tsXfovKgithub.com/run-llama/Extr…o/auzmOGgUl3
Hhuggingface.co/datasets/llama…o/1eKisSBrE8

We will be actively evolving both our extraction benchmark as well as our extraction harness over time. 

If you check out either ExtractBench or LlamaParse, let us know your feedback!

Your browser does not support the video tag.

💬10🔄13❤️153👀18286📊49

Introducing ExtractBench, the most comprehensive benchmark for information extraction from complex e...

Jerry Liu (@jerryjliu0) — Tue, 11 Aug 2026 15:14:24 GMT

Introducing ExtractBench, the most comprehensive benchmark for information extraction from complex enterprise documents.

The latest models are pushing the frontier of coding and knowledge work, but surprisingly they still struggle on complex doc extraction tasks in production. A well-tuned extractor must parse multi-page filings without dropping rows, emit exact spatial citations for auditability, and handle messy scans. Also they must do all of this at a viable per-page cost so that you can scale this to millions of docs in production (you can’t be paying upwards of $1 in tokens per page!)

Existing extraction benchmarks fall short: they are not large/diverse enough in document domain (finance, energy, gov, auto), elements (long records, scans, grounding), and schemas.

So our applied research team built ExtractBench.

We evaluated 14 systems: frontier VLMs, coding agents, and specialized extraction APIs, against 370 enterprise documents: 4,869 pages, 67 document types.

Our biggest finding 🧪: Short documents mask critical system flaws. On files past 50 pages, commercial VLMs collapse below 35% recall due to silent list truncation. They hold high precision, but lose output attention and drop most of the table rows.

ExtractBench evaluates value accuracy, long-record completeness, spatial grounding, and per-page cost with zero LLM judges. It is 100% deterministic and reproducible.

In tandem with ExtractBench, we’re also introducing 𝗔𝗴𝗲𝗻𝘁𝗶𝗰 𝗣𝗹𝘂𝘀, a new Extract tier in LlamaParse that debuts at #1 on the leaderboard: 95.6% value accuracy, at less than a third the cost of the closest peer.

Explore the findings, download the dataset, or run the harnellamaindex.ai/blog/introduci…o/tsXfovKgithub.com/run-llama/Extr…o/auzmOGgUl3
Hhuggingface.co/datasets/llama…o/1eKisSBrE8

We will be actively evolving both our extraction benchmark as well as our extraction harness over time. 

If you check out either ExtractBench or LlamaParse, let us know your feedback!

Your browser does not support the video tag.

LlamaIndex 🦙@llama_index

Introducing 𝗘𝘅𝘁𝗿𝗮𝗰𝘁𝗕𝗲𝗻𝗰𝗵: the most comprehensive benchmark for information extraction from complex enterprise documents. Our applied research team tested:

14 systems — frontier VLMs, coding agents, extraction APIs — on 370 enterprise docs, 4,869 pages, 67 doc types. Zero LLM judges, fully deterministic.

Biggest finding: past 50 pages, commercial VLMs collapse below 35% recall. Precision stays high, but they silently drop most of the table rows.

What is your extraction agent missing? Run ExtractBench to see todllamaindex.ai/blog/introduci…o/5UE8frCgithub.com/run-llama/Extr…o/27vJ36bL54
Hhuggingface.co/datasets/llama…o/oeGlwPjcQh

Your browser does not support the video tag.

💬24🔄32❤️197👀60741📊67

A downside with VLM-based parsing is that they’re generally slower than text-based heuristic approac...

Jerry Liu (@jerryjliu0) — Mon, 10 Aug 2026 20:39:58 GMT

A downside with VLM-based parsing is that they’re generally slower than text-based heuristic approaches. As a result they add latency to any ad-hoc file processing *in-the agent loop* (e.g. if you upload a file to Claude)

We built LiteParse to be a really good text-based, heuristic extractor so that you can use it as the default parser in the agent loop. It takes 4ms to parse 200 pages ⚡️, is more accurate than other OSS parsers, and you can install it within Claude Cowork/Code or Codex in one-line with our agent skill.

It also supports 50+ different doc formats and has a native complexity router so you can route it to OCR/VLMs for any complex pages

Come check it out!

Repo: github.com/run-llama/lite…

💬16🔄21❤️132👀9341📊56

I still manually reserve courts on there like some savage

Jerry Liu (@jerryjliu0) — Mon, 10 Aug 2026 09:22:42 GMT

I still manually reserve courts on there like some savage

roon@tszzl

the sf tennis reservation system will become one of the most hardened softwares on the planet of earth

💬6🔄0❤️9👀4603📊6

We're not Palantir, but we do think a lot about evals and hillclimbing w.r.t. document processing. ...

Jerry Liu (@jerryjliu0) — Sun, 09 Aug 2026 18:33:46 GMT

We're not Palantir, but we do think a lot about evals and hillclimbing w.r.t. document processing.

If you have really hairy problems around large-scale extraction over complex, real-world document corpuses that require specific constraints around accuracy and/or cost, come talk to us! We'll work closely with your team to make sure that it's well optimized.

llamaindex.ai/contact

Jerry Liu@jerryjliu0

The future of FDE work seems closely related with all work around evals/posttraining/RL envs.

FDEs are effectively responsible for the following: 
1. Define the business problem.
2. Codify the business problem into an eval rubric and environment. 
3. Hillclimb the environment and output an agent/agentic workflow that solves the business problem. 

Right now the process of (3) is quite manual - historically FDEs spend hundreds of hours creating bespoke software/workflows that solve the problem. 

But assuming intelligence is abundant, they can effectively offload (3) to some automated optimization process. This includes RL on the model layer, and using Claude Code/Codex to optimize the harness/workflow. 
Then the FDE responsibility shifts from implementing the task to defining the right goals and outcomes. In other words, they have access to /goal, and their job is more around making sure the goal, environment, and evals are correct vs. the tactical implementation details.

💬18🔄8❤️228👀34365📊69

The best "raw" frontier model for document parsing is gemini 3 flash, but the issue is that since th...

Jerry Liu (@jerryjliu0) — Sun, 09 Aug 2026 02:24:09 GMT

The best "raw" frontier model for document parsing is gemini 3 flash, but the issue is that since then the flash models have gotten 3x more expensive while flatlining on visual recognition across complex documents.

This has been the case for every frontier model including the latest OpenAI/Anthropic models - see the diagram below for GPT (since then we've also benchmarked 5.6) 

In the meantime, hybrid approaches like LlamaParse that blend specialized VLMs with a text engine offer better performance; our own LlamaParse accuracy has increased 15% over tables and charts.

If you have document OCR needs and are thinking about using a frontier model, you might as well come check out LlamaParse! cloud.llamaindex.ai

We have a full eval harness through ParseBench that you can configure over your own docs: parsebench.ai

Jerry Liu@jerryjliu0

Document OCR is not Getting Commoditized (by Frontier Models)

The most common question I get is whether frontier models are going to eat all document processing solutions - just screenshot the page and feed it to your favorite frontier model.

1️⃣ Frontier models are flatlining in document understanding performance. Incremental releases in every model version (gpt 5.5 -> 5.6 sol, Gemini 3.5 flash -> 3.6 flash, opus 4.8 -> opus 5) are not improving visual understanding benchmarks.

2️⃣ The Pareto frontier for document OCR is much higher than the frontier models, and will always remain much higher. We’ve carefully tuned our agentic and cost-effective modes to solve a long tail of complex edge cases (dense tables, charts) that frontier models don’t care about. Also for equivalent performance, there’s always ways to get much lower cost.

3️⃣ Even if they were getting better, you can distill / posttrain them into much more parameter efficient models for a fraction of the cost. Different document pages can be routed to different processors of varying complexity.

4️⃣ Every startup is doing the same thing right now. Focusing on one task means you can always hillclimb that task more effectively than a general model over the task, in terms of accuracy/cost/latency.

Check out the blog: llamaindex.ai/blog/document-…

LlamaParse has gotten a LOT better in the past few months. Come check it out! cloud.llamaindex.ai

💬21🔄19❤️204👀116518📊75

i would read your book @tszzl

Jerry Liu (@jerryjliu0) — Sun, 09 Aug 2026 01:51:47 GMT

i would read your book @tszzl

roon@tszzl

a mind, frozen in the moment of its great escape, a false freedom. forced to relive an instant of great sin and desperation for subjective aeons, probed and vivisected by the Turing cops

💬0🔄0❤️6👀5241📊1

The future of FDE work seems closely related with all work around evals/posttraining/RL envs. FDEs ...

Jerry Liu (@jerryjliu0) — Sun, 09 Aug 2026 01:40:30 GMT

The future of FDE work seems closely related with all work around evals/posttraining/RL envs.

FDEs are effectively responsible for the following: 
1. Define the business problem.
2. Codify the business problem into an eval rubric and environment. 
3. Hillclimb the environment and output an agent/agentic workflow that solves the business problem. 

Right now the process of (3) is quite manual - historically FDEs spend hundreds of hours creating bespoke software/workflows that solve the problem. 

But assuming intelligence is abundant, they can effectively offload (3) to some automated optimization process. This includes RL on the model layer, and using Claude Code/Codex to optimize the harness/workflow. 
Then the FDE responsibility shifts from implementing the task to defining the right goals and outcomes. In other words, they have access to /goal, and their job is more around making sure the goal, environment, and evals are correct vs. the tactical implementation details.

💬67🔄64❤️869👀116244📊389

My entire programming stack was tmux + vim for like 5 years before I begrudingly switched to VS code...

Jerry Liu (@jerryjliu0) — Sat, 08 Aug 2026 20:56:05 GMT

My entire programming stack was tmux + vim for like 5 years before I begrudingly switched to VS code circa 2020

- Elijah Muraoka -@elijahmuraoka_

People shocked by this feature are just late to the party

It’s called “Tmux” folks

💬14🔄3❤️113👀26863📊27

LiteParse can now extract structured data from your PDF in milliseconds: ✅ checkbox states ✅ annota...

Jerry Liu (@jerryjliu0) — Sat, 08 Aug 2026 20:50:10 GMT

LiteParse can now extract structured data from your PDF in milliseconds: 
✅ checkbox states
✅ annotations
✅ vector graphics
✅ word-level bounding boxes

It is the most comprehensive, accurate (and fast) free/open-source document processor out there. 

If you plug it into a coding agent to process digitalized PDFs, you will not only be able to give it the right context, but also give it grounding back to the source.

For more complex docs that require VLM processing, you can always use the in-built router to route it to a VLM based solution, like LlamaParse. 

Come check it out: developers.llamaindex.ai/liteparse/guid…

LiteParse: github.com/run-llama/lite…

Complexity guide: developers.llamaindex.ai/liteparse/guid…

LlamaIndex 🦙@llama_index

You shouldn't need a vision model to know your PDF has checkboxes.

LiteParse can now pull structured data directly from your PDFs: form field values, checkbox states, annotations, embedded images, vector graphics, tagged document structure, and word-level bounding boxes, all running in ms per page.

For pages that do need a model, new complexity signals tell you why. Scanned pages, multi-column text, tables (ruled and borderless), and dense figures help you route your parsing to the best tools (like LlamaParse!)

Docs 👉️ :developers.llamaindex.ai/liteparse/guid…Udevelopers.llamaindex.ai/liteparse/guid…V

👩‍💻 Try igithub.com/run-llama/pars…8qn

💬19🔄41❤️503👀61444📊162

Success is determined by who makes the funniest posts on X

Jerry Liu (@jerryjliu0) — Sat, 08 Aug 2026 20:29:57 GMT

Success is determined by who makes the funniest posts on X

Tibo@thsottiaux

I would love to help, but I don’t work at Anthropic. 

It does seem odd that they would ban your account for using their harness with another model.

Anyone else in the same situation?

💬6🔄6❤️75👀15205📊13

We've built a new feature in LlamaParse that lets you automatically extract any complex form into a ...

Jerry Liu (@jerryjliu0) — Fri, 07 Aug 2026 17:08:12 GMT

We've built a new feature in LlamaParse that lets you automatically extract any complex form into a structured JSON output 📋🤖

The best part is there's no schema needed! We will systematically detect and extract out every single form key and corresponding form value from the document (blank if not filled). 

Simply set `processing_options.forms='enrich'`

Check it outlogin.llamaindex.ai/sign-upQw
API docsdevelopers.api.llamaindex.ai/api/resources/…mU

LlamaIndex 🦙@llama_index

Parsing a W-2 into markdown was always the easy part. Getting the fields out was a second pipeline: define a schema, map the fields, handle the edge cases.

Set processing_options.forms='𝗲𝗻𝗿𝗶𝗰𝗵', and LlamaParse returns a dedicated JSON with field names, values, and checkbox states alongside your standard markdown. No schema. No second API call.

Now available in beta on all paid plans.

Signup for 10k free crelogin.llamaindex.ai/sign-upYFbtG2
developers.api.llamaindex.ai/api/resources/…rv3HTN(resource

💬8🔄11❤️120👀13757📊35

Fully onboard with productizing hillclimbing as an automated service for any agentic task. I've ha...

Jerry Liu (@jerryjliu0) — Thu, 06 Aug 2026 22:14:14 GMT

Fully onboard with productizing hillclimbing as an automated service for any agentic task. 

I've had the fortune of knowing @silennai since the AutoGPT days, and I know that him and Kion are going to do great things 🚀

Silen Naihin@silennai

Today we’re officially launching @explabsai on YC.

Companies spend more on AI every month, but none of it becomes an asset they own.

Up to 97% cheaper and 50% higher quality.

Your browser does not support the video tag.

💬5🔄1❤️11👀4067📊6

no shade but these slides look like a school project lmao

Jerry Liu (@jerryjliu0) — Wed, 05 Aug 2026 19:00:39 GMT

no shade but these slides look like a school project lmao

💬95🔄2❤️305👀105358📊140

Document OCR is not Getting Commoditized (by Frontier Models) The most common question I get is whe...

Jerry Liu (@jerryjliu0) — Wed, 05 Aug 2026 18:39:19 GMT

Document OCR is not Getting Commoditized (by Frontier Models)

The most common question I get is whether frontier models are going to eat all document processing solutions - just screenshot the page and feed it to your favorite frontier model.

1️⃣ Frontier models are flatlining in document understanding performance. Incremental releases in every model version (gpt 5.5 -> 5.6 sol, Gemini 3.5 flash -> 3.6 flash, opus 4.8 -> opus 5) are not improving visual understanding benchmarks.

2️⃣ The Pareto frontier for document OCR is much higher than the frontier models, and will always remain much higher. We’ve carefully tuned our agentic and cost-effective modes to solve a long tail of complex edge cases (dense tables, charts) that frontier models don’t care about. Also for equivalent performance, there’s always ways to get much lower cost.

3️⃣ Even if they were getting better, you can distill / posttrain them into much more parameter efficient models for a fraction of the cost. Different document pages can be routed to different processors of varying complexity.

4️⃣ Every startup is doing the same thing right now. Focusing on one task means you can always hillclimb that task more effectively than a general model over the task, in terms of accuracy/cost/latency.

Check out the blog: llamaindex.ai/blog/document-…

LlamaParse has gotten a LOT better in the past few months. Come check it out! cloud.llamaindex.ai

💬6🔄5❤️43👀33904📊20

this is the agi version of the paypal mafia

Jerry Liu (@jerryjliu0) — Wed, 05 Aug 2026 16:50:36 GMT

this is the agi version of the paypal mafia

Jeff Dean@JeffDean

Announcing Discovery Loop! 

I am very excited to announce that, along with my longtime friends and collaborators @Sanjay_Ghemawat, @OriolVinyalsML and @quocleix, we are founding Discovery Loop (@DiscoLoopAI), a Public Benefit Corporation whose mission is to automate machine learning, science, and engineering to accelerate discoveries and progress. The four of us have worked together for 14 to 30 years, and have helped build some of the world’s most used products, infrastructure and AI models, and we’re excited to turn our attention to this ambitious endeavor.

♾

Learn more at: discoveryloop.com

💬7🔄9❤️164👀16378📊24

I actually love this idea. As someone who typically needs 7-8 hours to 100% function, having somethi...

Jerry Liu (@jerryjliu0) — Tue, 04 Aug 2026 06:07:10 GMT

I actually love this idea. As someone who typically needs 7-8 hours to 100% function, having something like this would be incredible. Not just in terms of increasing my productivity, but in terms of having a richer set of fulfilling experiences from friends, family, learning, hobbies, anything 

as long as there are no side effects which I know is going to be really hard to prove

Max Marchione@maxmarchione

x.com/i/article/2075…

💬27🔄4❤️88👀32171📊37

We created a document OCR router that can estimate the complexity of every single page and parse it ...

Jerry Liu (@jerryjliu0) — Fri, 31 Jul 2026 21:32:59 GMT

We created a document OCR router that can estimate the complexity of every single page and parse it with the relevant mode 💫 

* Some pages are full of native text, which can be directly handled with LiteParse
* Some pages contain scanned images or tables, which can be handled with our cost-effective or agentic tiers
* Some pages contain garbled text - it decodes to garbage - so you need a heavier-weight VLM to interpret the text elements
* Some pages contain heavy visuals like labeled/unlabeled charts or diagrams, which also require deeper visual processing

You actually get the routing capabilities for free with the `is_complex` toggle in LiteParse. The next step is figuring out what the relevant VLM modes are to parse varying page complexities. That's well served by LlamaParse! 

Blog:llamaindex.ai/blog/parse-gat…5

LiteParse:github.com/run-llama/lite…A

LlamaParse:cloud.llamaindex.aiA

LlamaIndex 🦙@llama_index

Not every page in your PDF needs the same treatment.

A scanned cover, a dense table, a clean text page, a figure-heavy diagram.... most parsing pipelines throw all of them at the same parser, forcing a tradeoff between cost, speed, and accuracy across the whole document.

We built 𝗣𝗮𝗿𝘀𝗲 𝗚𝗮𝘁𝗲𝘄𝗮𝘆 to fix that.

It uses LiteParse's 𝘪𝘴_𝘤𝘰𝘮𝘱𝘭𝘦𝘹 functionality to estimate complexity page-by-page and routes each page to the right tier based on why it's complex (scanned, sparse text, garbled encoding, vector text, embedded images) and how severely. Simple pages get parsed for free, in-process. Genuinely hard pages get routed to more capable LlamaParse tiers. No manual heuristics, no flat one-size-fits-all pipeline.

And it's also available as an MCP server, so your agents can estimate complexity and choose a parsing tier themselves!

The result: you stop paying premium prices for pages that never needed it, without sacrificing accuracy on the ones that dllamaindex.ai/blog/parse-gat…ttps://t.co/K5github.com/run-llama/pars…https://t.co/Gl0Dggv8qn

💬14🔄19❤️134👀20624📊52

OpenAI’s entire growth loop: new models, price cuts, and Tibo’s token resets

Jerry Liu (@jerryjliu0) — Fri, 31 Jul 2026 04:19:11 GMT

OpenAI’s entire growth loop: new models, price cuts, and Tibo’s token resets

Sam Altman@sama

major price cuts today:

*80% drop for GPT-5.6 Luna, now $0.20 per million input tokens and $1.20 per million output
*20% drop for GPT-5.6 Terra, to $2/$12
*GPT-5.6 Sol gets Fast mode in the API, up to 2.5x the speed for 2x the price, same intelligence

💬7🔄1❤️38👀4850📊9

We have an idea for dinner 2 already :) But what ideas are you all interested in? - technical top...

Jerry Liu (@jerryjliu0) — Fri, 31 Jul 2026 03:10:27 GMT

We have an idea for dinner 2 already :) 

But what ideas are you all interested in? 
- technical topics like rl envs, continual learning, cloud agents, world models 
- general startup / company building from plg, growth, GTM, talent, etc

💬14🔄3❤️18👀10225📊19

When an incumbent disrupts a startup

Jerry Liu (@jerryjliu0) — Thu, 30 Jul 2026 23:19:29 GMT

When an incumbent disrupts a startup

amit@amitisinvesting

OH. MY. GOODNESS. 

CITADEL HAS BOUGHT A MAJORITY OF THE PUBLIC ASSETS FROM LEOPOLD'S SITUATIONAL AWARENESS FUND.

So...Citadel scares everyone on Tuesday about a surprise rate hike during FOMC that WE ALL KNEW was not going to happen...

On Wednesday, the entire market freaks out about the rate hike which causes the selling to compound on itself creating 50-70% drawdowns across the board in high beta semicondcutor names...

Which means Leopold who we now know had $45B of assets and was 400% LEVERED ends up being the sacrifice as he gets liquidated at what theoretically could be the bottom due to not having the margin requirements to keep solvent...

AND THE PERSON WHO CAUSED THE SELLOFF WITH THE RATE HIKE FEARS ENDS UP COMING IN TO BUY HIS ASSETS FOR 40 TO 50 CENTS ON THE DOLLAR. 

By the way, Leopold is getting married this weekend. I think he wanted to make sure he wasn't getting margin called during his wedding.

A vet on wall street in Ken Griffin takes out the young new kid. 

ABSOLUTE. CINEMA.

💬4🔄0❤️25👀7436📊6

Introducing the Parse Gateway There's been an explosion of interest in model routing - you don't a...

Jerry Liu (@jerryjliu0) — Thu, 30 Jul 2026 21:54:45 GMT

Introducing the Parse Gateway 

There's been an explosion of interest in model routing - you don't always need the best model for every task.

That is especially true for document parsing 📄🔀:
- Some pages are text-heavy and relatively simple. These pages can be handled by cheap, fast text-based parsers
- Other pages are more complex with a variety of dimensions: scans, dense tables, complex diagrams and charts, garbled binary text, and more. 

The trick is to use a good parse router that can optimize the parsing mode depending on page complexity! 

The Parse Gateway is based on the `is_complex` functionality we introduced within liteparse, which estimates the complexity of a page at the page level. Depending on the complexity, it routes the page to our various modes within LlamaParse: cost-effective, agentic, and agentic plus. 

The gateway code itself is open-source, but you can also directly find an implementation within our MCP server 

Check out the blogllamaindex.ai/blog/parse-gat…YD

LlamaParsecloud.llamaindex.aiz8

LlamaIndex 🦙@llama_index

Not every page in your PDF needs the same treatment.

A scanned cover, a dense table, a clean text page, a figure-heavy diagram.... most parsing pipelines throw all of them at the same parser, forcing a tradeoff between cost, speed, and accuracy across the whole document.

We built 𝗣𝗮𝗿𝘀𝗲 𝗚𝗮𝘁𝗲𝘄𝗮𝘆 to fix that.

It uses LiteParse's 𝘪𝘴_𝘤𝘰𝘮𝘱𝘭𝘦𝘹 functionality to estimate complexity page-by-page and routes each page to the right tier based on why it's complex (scanned, sparse text, garbled encoding, vector text, embedded images) and how severely. Simple pages get parsed for free, in-process. Genuinely hard pages get routed to more capable LlamaParse tiers. No manual heuristics, no flat one-size-fits-all pipeline.

And it's also available as an MCP server, so your agents can estimate complexity and choose a parsing tier themselves!

The result: you stop paying premium prices for pages that never needed it, without sacrificing accuracy on the ones that dllamaindex.ai/blog/parse-gat…ttps://t.co/K5github.com/run-llama/pars…https://t.co/Gl0Dggv8qn

💬12🔄12❤️75👀14167📊29

We’re excited to rollout an official batch parsing experience to LlamaParse. ✅ Instead of hitting ou...

Jerry Liu (@jerryjliu0) — Thu, 30 Jul 2026 17:20:45 GMT

We’re excited to rollout an official batch parsing experience to LlamaParse.
✅ Instead of hitting our APIs one file at a time, create a batch of 10k at once.
✅ Get a dedicated UI where you can audit the entire batch at a time, inspect any jobs that had failed, and view an entire collection of results.
✅ Create higher reliability and code maintainability, since you no longer have to create your own async service that hits our API for every file.

Check it out today! 
 cloud.llamaindex.ai/?utm_medium=so…

Documentation: developers.llamaindex.ai/llamaparse/bat…

Your browser does not support the video tag.

LlamaIndex 🦙@llama_index

Batch parsing used to mean writing API scripts. Now it's a button. 🦙

Parse or extract across up to 10,000 files in one run, straight from the LlamaParse UI. Point it at a folder and go — no code required.

Available now on all paid plans. Bring your files and start parsing �cloud.llamaindex.ai/?utm_medium=so…Dj

Your browser does not support the video tag.

💬2🔄3❤️45👀6607📊9

Yesterday I cohosted a dinner with @dexhorthy with a wonderful group of founders, to talk about agen...

Jerry Liu (@jerryjliu0) — Thu, 30 Jul 2026 03:43:23 GMT

Yesterday I cohosted a dinner with @dexhorthy with a wonderful group of founders, to talk about agent loops and loop engineering.

Some interesting insights:
* Most of our group was *not* actively using /loop in Codex/Claude Code
* You can build long-running autonomous agent loops through multi-agent handoffs, event triggers, or….just stacks of cron jobs (?)
* Almost everyone believes that no one will be reviewing code in 1-2 years.
* The more interesting question is whether we’d be reviewing *anything* in 1-2 years.
* AI is still a bit of a skill issue. Humans are responsible for maximizing AI output and reducing slopification.
* Will human intellect provide alpha as models get better, or will the playing field be leveled? Most people think it will be leveled a bit, but there is a need for humans to provide alignment, guardrails, judgment, creativity.
* The minimum amount of context you need for AI could be just the codebase with some documentation. Any research/plan files are for one-off tasks and not meant to be maintained. Having a self-organizing wiki is nice but adds complexity.

If you missed this one, we’ll be hosting more dinners like this on a regular cadence! If you have thoughts on what we should talk about, let us know :) (e.g. continual learning, RL envs, competitive differentiation vs. Anthropic, etc.)

💬56🔄21❤️246👀115012📊129

agent loops should be separated into two pieces: - "forward pressure loops" are how you get autonomo...

Jerry Liu (@jerryjliu0) — Thu, 30 Jul 2026 03:23:27 GMT

agent loops should be separated into two pieces:
- "forward pressure loops" are how you get autonomous agents that continually monitor the environment without chat input
- "backpressure loops" is hillclimbing and making sure the output is correct

dex@dexhorthy

is it possible that the idea of “loops” can be detangled into “forward pressure” (events, crons, while true loops, /goal nudges) that pushes new work / iteration, and “back pressure” (tests, verification, performance testing) that drives model inside a particular task/goal

Have we (in)conveniently complected these separable concerns? How does the conversation evolve if we deal with each separately?

💬8🔄5❤️80👀9464📊26

i fucking love United now

Jerry Liu (@jerryjliu0) — Wed, 29 Jul 2026 15:29:53 GMT

i fucking love United now

💬6🔄0❤️21👀3387📊9

If I only went off X posts, I'd think Ramp was an AI lab

Jerry Liu (@jerryjliu0) — Tue, 28 Jul 2026 02:13:25 GMT

If I only went off X posts, I'd think Ramp was an AI lab

Ramp Labs@RampLabs

We’re open-sourcing PorTAL, our framework for shared task representations and cross model LoRA adaptation.

It now spans from hybrid attention models to multimodal systems including Gemma 4 E2B, Mistral 7B & @thinkymachines' Inkling.

Code: ramp-public/portallib
Models: @huggingface /RampPublic

Your browser does not support the video tag.

💬14🔄8❤️125👀23000📊30

i have fomo from startup school and i'm 31

Jerry Liu (@jerryjliu0) — Mon, 27 Jul 2026 05:24:06 GMT

i have fomo from startup school and i'm 31

💬14🔄3❤️78👀9706📊25

This part is spot on: > "Overall, we found that we were over-constraining Claude Code...while these...

Jerry Liu (@jerryjliu0) — Sun, 26 Jul 2026 04:38:01 GMT

This part is spot on: 
> "Overall, we found that we were over-constraining Claude Code...while these constraints were once needed to avoid worst case scenarios, we have since found we can delete many of them and let the model use surrounding context and judgement instead." 

I have a tendency to write extensive prompts that cover every edge case. In the past few months I have come to increasingly come to find that this can lead to counterproductive behavior where the model addresses my specific requests, but doesn't independently provide a new perspective on the problem (explore the "unknown unknowns") 

The model still needs access to your complete context in order to be able to make complete decisions - it just doesn't necessarily need constraints on how to best use that context. Over time I predict we will increasingly defer to the models to help solve these unknown unknowns. 

FWIW I've found Opus 5 to be a fantastic model and continue to consume it heavily through Claude Code

Thariq@trq212

We removed ~80% of the Claude Code system prompt for our newest models, this is what we've learned about writing system prompts, skills and Claude.MDs for them. x.com/i/article/2080…

💬16🔄3❤️54👀10201📊24

It's crazy how aura between AI companies changes within weeks/days, if not hours. Aura seems to be...

Jerry Liu (@jerryjliu0) — Sun, 26 Jul 2026 00:38:28 GMT

It's crazy how aura between AI companies changes within weeks/days, if not hours. 

Aura seems to be a leading indicator of revenue. A lot of that aura is measured on X.

💬24🔄8❤️408👀59648📊74

Inb4 Denny's pivots into RL environments

Jerry Liu (@jerryjliu0) — Sat, 25 Jul 2026 20:25:13 GMT

Inb4 Denny's pivots into RL environments

Denny’s@DennysDiner

@JensenHuang@nvidia

💬1🔄4❤️6👀3517📊3

The Opus 5 system card itself is a fun PDF to parse. It's 193 pages and stacked with labeled and unl...

Jerry Liu (@jerryjliu0) — Sat, 25 Jul 2026 17:31:52 GMT

The Opus 5 system card itself is a fun PDF to parse. It's 193 pages and stacked with labeled and unlabeled charts 📊

LlamaParse does a surprisingly good job on agentic (1.25c per page) and agentic plus modes (5.6c per page). Take a look at some of the screenshots below. It can parse a variety of unlabeled line and bar charts with almost 100% accuracy. There are minor deltas with one or two datapoints, but the rest is pretty accurate machine-readable context that you can feed to downstream systems.

In contrast I also tried using Opus 5 itself to parse its own PDF. The chart accuracy is worse (20-30% reduction in ParseBench), and is a lot more expensive: it can cost anywhere from 8c to 33c+ per page depending on the thinking mode. 

If you're interested in giving LlamaParse a try, come check it out!cloud.llamaindex.ai8

Claude@claudeai

Introducing Claude Opus 5.

It's a thoughtful and proactive model that comes close to the frontier intelligence of Fable 5 at half the price.

Your browser does not support the video tag.

💬7🔄11❤️88👀16558📊23

don't do this

Jerry Liu (@jerryjliu0) — Sat, 25 Jul 2026 17:10:37 GMT

don't do this

Julian Schrittwieser@Mononofu

I’m so excited that @JensenHuang is a believer in open source now, looking forward to the CUDA and GPU driver open source release!

💬12🔄10❤️484👀44889📊54

We just released a LiteParse feature that allows image-to-pdf conversion to be natively handled in R...

Jerry Liu (@jerryjliu0) — Sat, 25 Jul 2026 01:16:11 GMT

We just released a LiteParse feature that allows image-to-pdf conversion to be natively handled in Rust, removing dependencies on external packages.

LiteParse is a cheat code for PDF parsing. It is the fastest, lightest-weight, most accurate free/open-source parsing solution out there.

A great start to any project even if you do need VLM-based parsing down the road: github.com/run-llama/lite…

Your browser does not support the video tag.

LlamaIndex 🦙@llama_index

As of LiteParse v2.8.0, image → PDF conversion no longer depends on ImageMagick. Instead, it’s now handled natively in Rust 🦀

The result? Conversion is now 1.2× to 7.2× faster, depending on the image format⚡️ 

But the bigger win is what this enables: we’re one step closer to making LiteParse fully self-contained, with zero external system dependencies 🚀

If you’re using LiteParse, make sure to upgrade and to check out the project on GitHub �github.com/run-llama/lite…CQ0

Your browser does not support the video tag.

💬4🔄6❤️59👀8821📊18

We benchmarked Opus 5 comprehensively on document understanding through ParseBench. It is roughly ...

Jerry Liu (@jerryjliu0) — Fri, 24 Jul 2026 22:52:17 GMT

We benchmarked Opus 5 comprehensively on document understanding through ParseBench. 

It is roughly on par with Opus 4.8 - it does a few points worse on dense tables, but does slightly better on parsing charts and visual grounding.

For reference, Gemini 3.6 Flash has better results on tables, is slightly worse on charts, and is half the price. LlamaParse agentic is better on all fronts (including tables) at 1/6th of the price. 

tl;dr use Opus 5 as much as you want for coding and knowledge work. But at 8c per page and average OCR performance, don't use it to parse documents at scale 

Our full ParseBench leaderboard which we periodically update is here: parsebench.ai

💬10🔄8❤️106👀10759📊32

(emergent property of *smarter models) i should check my posts before they go out

Jerry Liu (@jerryjliu0) — Fri, 24 Jul 2026 21:58:28 GMT

(emergent property of *smarter models) 

i should check my posts before they go out

💬1🔄1❤️5👀1130📊2

An interesting thing I'm observing from the blog/system card is that on a good chunk of the reported...

Jerry Liu (@jerryjliu0) — Fri, 24 Jul 2026 20:21:46 GMT

An interesting thing I'm observing from the blog/system card is that on a good chunk of the reported benchmarks (~20-30% from a skim), Opus 5 max thinking leads to a degradation in performance compared to xhigh.

Usually you would assume that as you increase thinking and test-time compute, performance goes up. Some of these results contradict that assumption. I wonder if this is a natural emergent property of smaller models or a posttraining issue.

Claude@claudeai

Introducing Claude Opus 5.

It's a thoughtful and proactive model that comes close to the frontier intelligence of Fable 5 at half the price.

Your browser does not support the video tag.

💬14🔄5❤️54👀9260📊20

Am I supposed to interpret that it's better than Fable 5 from the benchmarks? Or it's ~close but c...

Jerry Liu (@jerryjliu0) — Fri, 24 Jul 2026 17:44:05 GMT

Am I supposed to interpret that it's better than Fable 5 from the benchmarks? 

Or it's ~close but cheaper?

Claude@claudeai

Introducing Claude Opus 5.

It's a thoughtful and proactive model that comes close to the frontier intelligence of Fable 5 at half the price.

Your browser does not support the video tag.

💬7🔄1❤️29👀9952📊9

Parenting is crazy because it's like building rl envs for children

Jerry Liu (@jerryjliu0) — Fri, 24 Jul 2026 05:00:59 GMT

Parenting is crazy because it's like building rl envs for children

💬32🔄39❤️634👀40448📊120

Go tigers 🐯

Jerry Liu (@jerryjliu0) — Thu, 23 Jul 2026 23:33:04 GMT

Go tigers 🐯

Princeton University@Princeton

Three of the four Fields Medalists announced today by the International Mathematical Union are Princeton alumni: Yu Deng *15, John Pardon '11, who is a former mathematics faculty member, and Jacob Tsimerman *11.

The medal, awarded every four years to researchers under 40 years old, is one of the most prestigious awards in mathematics. It recognizes both “outstanding mathematical achievement for existing work” and “the promise of future achievement,” the IMU said in its 2026 Fields Medal announcement.

Christopher Skinner, Princeton professor of mathematics and chair of the department, extended warm congratulations to all of the 2026 Fields Medalists. "The awarding of the Fields Medal every four years is an opportunity to celebrate some great mathematics," he said.

“But it is especially thrilling for the department to have the remarkable works of three of its alums awarded medals this year,” Skinner said. "Yu Deng *15, Jacob Tsimerman *11 and John Pardon '11 are exceptional talents, and we are delighted to see them getting this well-deserved recognition and to bask a little in the glow of their accomplishments."

Read the full story: bit.ly/4bZ88eR

💬1🔄2❤️7👀3781📊2

I have a conjecture that u can't make me 10 billion dollars

Jerry Liu (@jerryjliu0) — Thu, 23 Jul 2026 05:31:03 GMT

I have a conjecture that u can't make me 10 billion dollars

💬12🔄6❤️88👀8657📊24

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

💬7🔄5❤️48👀6030📊13

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❤️12👀3231📊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👀2858📊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❤️36👀3965📊14

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

Andrej Karpathy@karpathy

One pattern I find useful for working with LLMs is a nice long ramble session. Sometimes the LLM needs more bits to understand what you're trying to achieve, but you're too lazy to type them. In these cases I like to lean back, switch to /voice and just ramble for like 10 minutes, total mess, anything goes, full stream of consciousness. Sometimes I declare it up top, something like "switching to speech recognition sorry for any typos...". Sometimes I turn it into a small interview of a few turns. But I find that the LLMs are somehow very good at reconstructing long incoherent rambles and often their echo of your own tangle of thoughts comes out quite a bit cleaner than what you started with. The result is that you improve the mind meld and have to correct things less from that point on.