elvis(@omarsar0)

The hard part of multi-agent systems is getting agents to stay quiet. Put five agents on one task, ...

elvis (@omarsar0) — Thu, 23 Jul 2026 17:14:07 GMT

The hard part of multi-agent systems is getting agents to stay quiet.

Put five agents on one task, and they duplicate work and burn tokens talking to each other.

Offloop trained a dispatcher model called D1 that decides which agent moves next and when the right move is to do nothing. They achieve state-of-the-art performance on GDPval at a fraction of the usual cost.

You can bring your own AI subscription. offloop.org

Your browser does not support the video tag.

🔗 View on Twitter

Offloop@Offloop

Introducing Offloop!

We're a team of four. Today our multi-agent harness hit state of the art on GDPval, ahead of Claude code and Codex across jobs that pay $2.4 trillion a year in the US.

Offloop gives every knowledge worker what the Fortune 500 spends billions on: a high-performing agent army that runs itself and grows the business.

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬5🔄0❤️11👀2609📊6

⚡ Powered by xgo.ing

Peter knows something, of course. But all these loops and graphs are just terms that will consolidat...

elvis (@omarsar0) — Thu, 23 Jul 2026 16:34:59 GMT

Peter knows something, of course. But all these loops and graphs are just terms that will consolidate into one interface that allows the complex agentic workflows and interactions that the task demands. And it should work dynamically and automatically. 

x.com/steipete/statu…

Peter Steinberger 🦞@steipete

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

🔗 View Quoted Tweet

💬0🔄0❤️3👀1126📊1

⚡ Powered by xgo.ing

It's clear to me that a lot of these ideas are very early. But having built my own agent orchestrato...

elvis (@omarsar0) — Thu, 23 Jul 2026 16:08:38 GMT

It's clear to me that a lot of these ideas are very early. But having built my own agent orchestrator has allowed me to experiment with agents in ways that are still hard to comprehend. I am still on the hunt for the most generalized form of an agent harness, but this might be close to it. I am doing intensive AI research in the background to benchmark these ideas, and it's a lot of work, but I am learning so much, and it's fun to share these ideas early here.

💬2🔄0❤️5👀1440📊2

⚡ Powered by xgo.ing

Dynamic workflows are a generalization of harnesses, automations, loops, routing, and graphs. It's...

elvis (@omarsar0) — Thu, 23 Jul 2026 16:04:30 GMT

Dynamic workflows are a generalization of harnesses,  automations, loops, routing, and graphs.

It's the most powerful feature I have built into my agent orchestrator.

Supports all kinds of patterns that leverage different agent backends (claude, codex, pi, hermes,...).

It's a meta-harness approach that unlocks new forms of test-time compute. 

Example of use cases it supports:

> LLM councils to get different perspectives from LLMs or plan more intensively

> Dynamically routing tasks to different agents based on needs (e.g., cost efficiency and optimal intelligence)

> Advisor/Judge + executor workflows and pretty much any complex graph-based pattern required by the task. I find it especially useful for long-running work and code reviewing.

> Agent teams that talk to each other if needed for the task. I like to use this for AI editing, artifact creation, and other creative tasks.

And I am sure it supports so many things that I haven't discovered yet. 

I got inspired by the dynamic workflow feature released by the Claude Code team. I had actually built it earlier this year but wanted to generalize it across different agent backends. I think this is going to become more popular in the coming days. I will share more of my findings soon.

Your browser does not support the video tag.

🔗 View on Twitter

💬10🔄3❤️47👀4535📊20

⚡ Powered by xgo.ing

Great paper on self-improving agent harnesses. (bookmark it) If you maintain a production agent ha...

elvis (@omarsar0) — Thu, 23 Jul 2026 14:20:01 GMT

Great paper on self-improving agent harnesses.

(bookmark it)

If you maintain a production agent harness, finding every file behind one behavior is often harder than writing the edit.

Harness Handbook builds a three-level map from runtime behaviors to source locations using static analysis and LLM-assisted structuring.

Its BGPD workflow guides coding agents from the system overview to relevant stages, functions, and files, then verifies every candidate against current source.

Across 60 modification requests on Codex and Terminus-2, handbook guidance raised planning win rates from 28.3% to 38.3% and from 26.7% to 45.6%.

Planner token use fell 12.7% and 8.6%.

File- and symbol-level F1 improved in all 24 comparisons against GPT-5.5 and Opus 4.8 reference plans. Complete localization misses fell by as much as 25.9 points.

This is a strong pattern for coding agents that need to evolve large harnesses without losing scattered or rarely executed behavior.

Paper: arxiv.org/abs/2607.13285

Learn to build effective AI agents in our academy: academy.dair.ai

💬16🔄24❤️112👀7406📊55

⚡ Powered by xgo.ing

Just implemented a new interface for our @dair_ai community. I first brainstormed a concept design...

elvis (@omarsar0) — Wed, 22 Jul 2026 21:18:37 GMT

Just implemented a new interface for our @dair_ai  community.

I first brainstormed a concept design with Fable 5 using an HTML artifact (attached image). Then I gave it to GPT-5.6-Sol to build it.

Artifacts work great as prompts. Combining models gives you efficiency gains.

💬6🔄1❤️30👀9399📊11

⚡ Powered by xgo.ing

In addition, something I have been building in my orchestrator system is to dynamically route to sev...

elvis (@omarsar0) — Wed, 22 Jul 2026 20:59:09 GMT

In addition, something I have been building in my orchestrator system is to dynamically route to several models at once when needed. Or split the task to be handled by different models. I will share something similar I have been working on sometime this week.

💬0🔄0❤️5👀1567📊1

⚡ Powered by xgo.ing

I will be doing a live session on this for our community. Join us here: https://t.co/aG0VzVZyhP

elvis (@omarsar0) — Wed, 22 Jul 2026 20:42:16 GMT

I will be doing a live session on this for our community. Join us here: academy.dair.ai/events/czm9avw…

💬0🔄1❤️5👀1466📊2

⚡ Powered by xgo.ing

It sucks that with most coding agents you have to call out the skill every single time. The agent ...

elvis (@omarsar0) — Wed, 22 Jul 2026 18:08:11 GMT

It sucks that with most coding agents you have to call out the skill every single time. 

The agent should just know when and what set of skills to use. @boltdotnew Skills solves this with skill auto-matching.

This compounds nicely when you build with agents as a team.

bolt.new@boltdotnew

There's a whole new way to use skills in Bolt.

Every skill your teammates build is now yours too. And they stack: one prompt triggers all of them, automatically.

Here's how it works 🧵

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬3🔄2❤️18👀9408📊5

⚡ Powered by xgo.ing

Karpathy's point about long voice rambles goes further when you add more modalities. My favorite ...

elvis (@omarsar0) — Wed, 22 Jul 2026 17:33:37 GMT

Karpathy's point about long voice rambles goes further when you add more modalities.  

My favorite way to prompt agents lately is a bigger unit I've been calling a task.  

A task bundles a long voice note, the current screen, annotations, and exact text into a single turn. The agent reconstructs my intent from all of those signals, so most of the correction loops disappear and I can hand off larger pieces of work.

elvis@omarsar0

x.com/i/article/2079…

🔗 View Quoted Tweet

💬17🔄17❤️119👀14963📊39

⚡ Powered by xgo.ing

https://t.co/juAGYwoKIE

elvis (@omarsar0) — Wed, 22 Jul 2026 17:30:44 GMT

x.com/i/article/2079…

💬3🔄14❤️57👀22986📊30

⚡ Powered by xgo.ing

Are structured outputs in agents always good? This paper suggests that you might have to take a clo...

elvis (@omarsar0) — Wed, 22 Jul 2026 16:01:06 GMT

Are structured outputs in agents always good?

This paper suggests that you might have to take a closer look.

Your product's structured output surface is measurably more homogeneous than the chat surface you benchmarked it on.

Teams treat structured output as a formatting choice with no effect on content. This new work from Cornell tests that assumption across 44 models and finds the format itself reshapes the answers.

On an open-ended pick-a-word prompt, simply requesting JSON pushes the modal answer from 41% to 64% of responses and cuts distinct answers from 52 to 36.

The collapse shows up for JSON and XML, the formats models were post-trained to emit, and disappears for YAML and CSV. Enforcing a schema at the decoder compresses nothing further, which points the narrowing back to how the model answers the register itself.

Why does it matter?

Any pipeline leaning on JSON mode for variety is sampling from a smaller pool than it thinks.

Paper. arxiv.org/abs/2607.18476

Learn to build effective AI agents in our academy. academy.dair.ai

💬18🔄13❤️92👀11133📊41

⚡ Powered by xgo.ing

It looks like it only works for Claude Cowork. That's really annoying. I think it's really worth it ...

elvis (@omarsar0) — Wed, 22 Jul 2026 15:48:39 GMT

It looks like it only works for Claude Cowork. That's really annoying. I think it's really worth it to build your own orchestrator tool and build these features out yourself. This is the future of how to work with agents, so you can't just be waiting for someone to build this out for you. There are many interesting variants of this as well.

💬2🔄1❤️4👀1446📊3

⚡ Powered by xgo.ing

I agree with what this AI paper suggests. Self-improving agents should evolve their benchmarks too....

elvis (@omarsar0) — Wed, 22 Jul 2026 14:05:02 GMT

I agree with what this AI paper suggests.

Self-improving agents should evolve their benchmarks too.

(bookmark it)

Self-improving agents are one of the most important directions in AI right now, and most of them optimize against a fixed benchmark. This paper argues the benchmark should evolve alongside the agent.

The setup is a self-evolving Lean proof agent. A small trusted runtime wraps a fully mutable workspace of workflow, prompts, and tools, so the agent can rewrite how it decomposes proofs, uses compiler feedback, and repairs failures.

What makes it different is coevolution. Between generations the champion revises the active task distribution through a mastery-throttled curriculum that adds harder proof obligations only after the current level is mastered.

All of it stays inside a Lean-grounded loop, so a success counts only when the behavior yields verified proofs under a trusted snapshot. Over 15 generations the best coevolving agent reaches 45.1% held-out solve rate on miniF2F, versus 12.7% for the seed and 32.0% for the best fixed-benchmark agent.

Grounding every reward in a formal verifier is how you get self-modification without reward hacking, and coevolving the curriculum pushes the ceiling higher.

Paper: arxiv.org/abs/2607.17352

Learn to build effective AI agents in our academy: academy.dair.ai

💬12🔄16❤️96👀11781📊42

⚡ Powered by xgo.ing

I was a bit too early on this. But I highly recommend reading it. It essentially packages two id...

elvis (@omarsar0) — Wed, 22 Jul 2026 00:33:03 GMT

I was a bit too early on this. 

But I highly recommend reading it. 

It essentially packages two ideas mentioned by both Claude (recording skills) and Andrej Karpathy (rich multimodal prompting) today. 

Lots of people seem to be picking this up now.

elvis@omarsar0

x.com/i/article/2073…

🔗 View Quoted Tweet

💬5🔄6❤️81👀20494📊27

⚡ Powered by xgo.ing

Progressive disclosure in agents doesn't scale. And its benefits seems agent harness dependent. (b...

elvis (@omarsar0) — Wed, 22 Jul 2026 00:00:09 GMT

Progressive disclosure in agents doesn't scale.

And its benefits seems agent harness dependent.

(bookmark this one)

Finally there is a proper study on using agent skills and the effect of progressive disclosure.

Progressive disclosure is the agent skills pattern where you hand an agent a document path and let it decide what to read, from a short description down to specific passages.

Practitioners adopted it fast for book-length tasks, purely on vibes.

Researchers ran it across three agent harnesses and three model families on InfiniteBench.

On a single book:

> the gain is harness-dependent,
> large when the agent navigates raw documents poorly,
> near zero when a strong harness already retrieves on its own.

Scale to many books and raw navigation falls apart while one level of disclosure pulls ahead. A second routing level never helps and sometimes breaks accuracy.

It feels like progressive disclosure buys context, but not intelligence. It is redundant while a strong agent can find the passage itself, and decisive once the corpus is too large to read.

Paper: arxiv.org/abs/2607.17598

Learn to build effective AI agents in our academy: academy.dair.ai

💬20🔄17❤️97👀11246📊44

⚡ Powered by xgo.ing

Love it how Jack keeps pushing for open-source AI. First Goose, now Buzz (looks like a great alterna...

elvis (@omarsar0) — Tue, 21 Jul 2026 23:30:11 GMT

Love it how Jack keeps pushing for open-source AI. First Goose, now Buzz (looks like a great alternative to Claude Tag in Slack). Open-source, multiplayer, and many other goodies. You see, Jack understands how important sovereignty and owning your intelligence is. Worth a try.

jack@jack

we're launching BUZZ!

a new groupchat platform for teams of people and agents of all sizes, built to reduce our dependency on slack and github. model-agnostic, decentralized, self-sovereign, and open source. 🐝
buzz.xyzo

🔗 View Quoted Tweet

💬4🔄2❤️21👀9932📊6

⚡ Powered by xgo.ing

Open weight models galore! Laguna S 2.1 went from the start of training to launch in under 9 weeks!...

elvis (@omarsar0) — Tue, 21 Jul 2026 22:58:44 GMT

Open weight models galore!

Laguna S 2.1 went from the start of training to launch in under 9 weeks! 

Custom & open models are putting a lot of pressure on the frontier closed and general models. No wonder they want to ban them. Diverse intelligence is good for the market.

Poolside@poolsideai

Today we're releasing Laguna S 2.1, our most capable model to date.

It's a 118B total parameter Mixture-of-Experts model with 8B activated per token, a context window of up to 1M tokens, and thinking and no-thinking modes.

Capable enough to hold its own against models many times its size. Small enough to run on a single @NVIDIAAI DGX Spark.

Laguna S 2.1 is fully open under OpenMDW-1.1, with weights available today on @huggingfacepoolside.ai/blog/introduci…

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬1🔄8❤️49👀9026📊9

⚡ Powered by xgo.ing

Thesean just launched Ship in beta, an endpoint that cuts your LLM bill in half. You keep the model...

elvis (@omarsar0) — Tue, 21 Jul 2026 19:10:17 GMT

Thesean just launched Ship in beta, an endpoint that cuts your LLM bill in half.

You keep the model name you already use. Change model="original" to model="ship-like/original" and every request costs a fixed 50% less.

Ship is built to preserve your prompts, output formats, and tool behavior.

Behind the scenes, Ship picks the most efficient way to run each request while matching the quality of the model you named. If a request costs more to execute than you paid, Thesean covers the difference.

The savings can go two ways. Cut the bill, or keep the same budget and run twice as many agents, retries, and verification passes to further improve output quality.

Martian@withmartian

Announcing Ship: an endpoint with the highest intelligence per dollar of any frontier model.

Today, Ship makes using Opus and GPT 5.6 Sol 50% cheaper guaranteed by a quality SLA.

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬4🔄3❤️17👀9533📊6

⚡ Powered by xgo.ing

Not because Andrej is saying it but I think voice is goated. And you can mix it with other modaliti...

elvis (@omarsar0) — Tue, 21 Jul 2026 18:33:50 GMT

Not because Andrej is saying it but I think voice is goated.

And you can mix it with other modalities for even richer prompting.

I recorded a session a few weeks back to demo the power of multimodal prompting x.com/omarsar0/statu…

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.

🔗 View Quoted Tweet

💬10🔄5❤️77👀20191📊21

⚡ Powered by xgo.ing

Great to see more people finding success with multimodal prompting

elvis (@omarsar0) — Tue, 21 Jul 2026 18:28:07 GMT

Great to see more people finding success with multimodal prompting

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.

🔗 View Quoted Tweet

💬1🔄1❤️6👀1818📊2

⚡ Powered by xgo.ing

This is a neat feature. I wrote an article a few weeks back about how I built this into my agent o...

elvis (@omarsar0) — Tue, 21 Jul 2026 16:36:59 GMT

This is a neat feature. 

I wrote an article a few weeks back about how I built this into my agent orchestrator:
x.com/omarsar0/statu…

But I made it multimodal from the ground up. Text, screenshots, audio, video, and annotations can all be wired up as a reusable skill. Magical!

Claude@claudeai

New in Claude Cowork: teach Claude a skill.

Record your screen while you do a task, talk through it as you go, and Claude turns it into a skill it can run again. Find it under Record a skill in the + menu of the Claude desktop app. 

Available on Pro, Max, and Team plans.

🔗 View Quoted Tweet

💬11🔄9❤️126👀28794📊30

⚡ Powered by xgo.ing

This is a radical improvement in how to use AI. @magicX_ai's AI Autocomplete predicts the actions a...

elvis (@omarsar0) — Tue, 21 Jul 2026 16:25:30 GMT

This is a radical improvement in how to use AI.

@magicX_ai's AI Autocomplete predicts the actions a user could take from any text box and surfaces them in real time as they type, with responses in 200ms. Any product can drop it in through an SDK.

Search captures intent after you already know what you want. AI Autocomplete reads it while you are still forming the thought. 

This opens great experiences on AI-powered products like search and customer service. Worth a look.

brad@bradkowalk

We just solved the biggest UX problem in AI.

Today, we're releasing AI Autocomplete, a SDK that gives every text box a brain.

It boosts your conversions by 50%+ by guiding your users as they type with what your product can do, and 500+ companies have already signed up.

So check it out if you have a search box or assistant.

Add it in under 5 mins ↓

ai-autocomplete.com

(or comment your product name, and we'll reply with a demo of it working on your product)

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬3🔄5❤️42👀10151📊12

⚡ Powered by xgo.ing

Octen's web search latency is pretty insane. Deep research tools like OpenAI, Gemini, Grok, and Pe...

elvis (@omarsar0) — Tue, 21 Jul 2026 16:05:41 GMT

Octen's web search latency is pretty insane. 

Deep research tools like OpenAI, Gemini, Grok, and Perplexity can take up to an hour to finish tasks.

Octen can return a full source-backed report in under 3 minutes, and scores 10 to 17 points higher than all four on DeepResearch Bench.

Landing both speed and accuracy is rare in web search. Highly recommend AI devs check it out if building with web search.

Kuan Zou@KZouAPT

x.com/i/article/2079…

🔗 View Quoted Tweet

💬3🔄4❤️24👀9682📊8

⚡ Powered by xgo.ing

Don’t sleep on these orchestration models. Sakana just announced Fugu-Cyber which achieves SoTA on r...

elvis (@omarsar0) — Tue, 21 Jul 2026 01:57:42 GMT

Don’t sleep on these orchestration models. Sakana just announced Fugu-Cyber which achieves SoTA on real-world security benchmarks, matching frontier models like GPT-5.5-Cyber and Mythos Preview.

Sakana AI@SakanaAILabs

Introducing Fugu-Cyber: an update to our Fugu orchestration model.

It achieves state-of-the-art performance on real-world security benchmarks, matching cyber-focused frontier models like GPT-5.5-Cyber and Mythos Preview.

sakana.ai/fugu-cyber-rel… 🐡

🔗 View Quoted Tweet

💬10🔄10❤️57👀11986📊15

⚡ Powered by xgo.ing

Huge launch from @tryramp. Different steps in an agent workflow can use different models. This can ...

elvis (@omarsar0) — Mon, 20 Jul 2026 22:09:00 GMT

Huge launch from @tryramp.

Different steps in an agent workflow can use different models. This can help reduce costs significantly without sacrificing performance. 

Model routing will become a core part of the stack.

Now available in one OpenAI-compatible endpoint.

Veeral Patel@vral

Today we’re launching Ramp Router.

3 years ago, we built an internal LLM router at @tryramp that powers AI products for 70,000 customers.

Back then it was mostly about saving money.

Now it feels obvious: the best model changes constantly. GPT, Claude, Gemini, Grok, Qwen, DeepSeek, Kimi, GLM - prices and capabilities move every week.

So we’re opening up access to everyone.

One OpenAI-compatible endpoint. The right model for every request. Lower cost without rewriting your app.

Reserve access to use it.

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬8🔄3❤️21👀10782📊9

⚡ Powered by xgo.ing

Something really neat about this work that might not get picked up easily: for post-training, you ma...

elvis (@omarsar0) — Mon, 20 Jul 2026 19:38:30 GMT

Something really neat about this work that might not get picked up easily: for post-training, you may not need long-horizon rollouts to train long-horizon agents. You could train on short tasks where rewards are cheap and verification is easy, and let the harness carry it 8-32x further.

💬0🔄0❤️4👀1767📊1

⚡ Powered by xgo.ing

Wanted to add that I care about this problem. As an independent researcher, I've been working hard a...

elvis (@omarsar0) — Mon, 20 Jul 2026 18:13:20 GMT

Wanted to add that I care about this problem. As an independent researcher, I've been working hard at @dair_ai on AI-powered research tools that help with things like forming hypotheses, finding interesting research directions, formulating strong research questions, compiling literature, reproducing works with coding agents, and much more. We are launching several tools in the coming months to help with all of this. Stay tuned.

💬1🔄2❤️9👀1835📊3

⚡ Powered by xgo.ing

Highly recommended. I've often claimed there's huge alpha in building agent harnesses. Turns out ...

elvis (@omarsar0) — Mon, 20 Jul 2026 16:56:31 GMT

Highly recommended.

I've often claimed there's huge alpha in building agent harnesses. 

Turns out harnesses are compositional generalizers. The RLM harness is an instance of this. This could lead to interesting and efficient new ways to scale generalization in models.

alex zhang@a1zhang

Transformers struggle to generalize to tasks they were not explicitly trained on. Instead, we propose in 2026 that it is the job of the harness to generalize through composition.

We observe a powerful property when training RLMs: for tasks with shared structure that look different, the root model naturally learns the same trajectory, meaning it views the two task trajectories as the same! In other words, the Transformer does not need additional generalization capabilities to transfer capabilities from one task to the other, the harness induces it.

We find that well-designed harnesses form a quotient set over task trajectories, meaning their individual LLM calls can see structurally “similar” tasks as near-identical, token-for-token! Harnesses can effectively generalize for the Transformer during training, without relying on any intrinsic generalization capability from the model.

For example, RLMs can see problems of different lengths as the same: we show that RLMs can train exclusively on short tasks, and fully generalize to similar but unseen tasks 8-32x longer because it produces near identical trajectories for both.

Taking this further, we show that tasks across different domains (e.g. math solutions vs. essay writing) that share a decomposition strategy exhibit the same generalization effect. RLMs can train on the problem of finding which essays belong to the same author and improve performance on finding math problems that share similar solutions.

The full blogpost, experiments, and discussion are in the thread below.

🔗 View Quoted Tweet

💬16🔄31❤️162👀30684📊50

⚡ Powered by xgo.ing

// Global Workspace in LLMs // arXiv paper for the popular J-space work from Anthropic. (bookmark ...

elvis (@omarsar0) — Mon, 20 Jul 2026 16:01:05 GMT

// Global Workspace in LLMs //

arXiv paper for the popular J-space work from Anthropic.

(bookmark it)

The short recap:

If you build on chain-of-thought or steering vectors, this work provides a mechanistic account of when verbalized reasoning is load-bearing and when it is narration after the fact.

The representations a model can put into words behave like a shared global workspace, a bandwidth-limited channel that broadcasts a small set of features to the rest of the network and steers what it does next.

Paper: arxiv.org/abs/2607.15495

Learn to build effective AI agents in our academy: academy.dair.ai

💬22🔄14❤️81👀11725📊43

⚡ Powered by xgo.ing

A lot of young researchers will look at this and feel lost and hopeless. Some might be thinking: “I...

elvis (@omarsar0) — Mon, 20 Jul 2026 14:30:13 GMT

A lot of young researchers will look at this and feel lost and hopeless.

Some might be thinking: “If AI can do this, what’s the point of what I am doing?”

I assure you that the reality is the exact opposite.

The opportunities AI unlocks are endless. Don’t let anyone or any company convince you otherwise.

AI is a brilliant tool. Use it to explore more and go deeper. Explore ideas in parallel. Use it to explore ideas that extend to multiple fields; these are extremely hard to do and extremely time consuming. 

Spend time brainstorming with AI. Use it to pose interesting and unique research questions. Brilliant research is highly dependent on the quality of research questions. Use it to find better ways to communicate the science and discoveries.

If I was starting my PhD all over again, I would certainly be doing that. 

It’s important to take the time to assess how you can leverage such tools. Now is a good time to do so. It’s inevitable that AI will lead to more of these findings but your intuition and research pedigree can accelerate what’s possible with it.

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)

🔗 View Quoted Tweet

💬18🔄27❤️204👀26571📊41

⚡ Powered by xgo.ing

This is not something that anyone can just one shot with an LLM. In my experience with LLMs, it does...

elvis (@omarsar0) — Mon, 20 Jul 2026 14:15:48 GMT

This is not something that anyone can just one shot with an LLM. In my experience with LLMs, it does require strong knowledge and intuition about the problem you are trying to solve. Really hope there is a report on the methodology.

💬1🔄0❤️3👀1421📊1

⚡ Powered by xgo.ing

That being said, I think it’s important to understand the methodology and the LLM traces behind such...

elvis (@omarsar0) — Mon, 20 Jul 2026 14:10:06 GMT

That being said, I think it’s important to understand the methodology and the LLM traces behind such finding. Hope there is a more detailed report about it soon.

Другая планета@nihilunbounded

Thank you! Vitushkin's paper makes Fable's example much more paletable. It's a shame Fable (or Levent, but that seems unlikely) didn't cite it. I noticed the xy + 1, xz terms in Fable's solution but its significance wasn't quite so transparent... nice find.

🔗 View Quoted Tweet

💬2🔄0❤️1👀2818📊2

⚡ Powered by xgo.ing

What!? It looks correct. I am convinced we are not pushing these LLMs enough. Benchmarks are simpl...

elvis (@omarsar0) — Mon, 20 Jul 2026 12:56:13 GMT

What!? It looks correct. 

I am convinced we are not pushing these LLMs enough. Benchmarks are simply not enough. That’s exciting.

Now can you please remove all the unnecessary guardrails from Fable so more of us can try solving extremely hard problems in our respective fields?

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)

🔗 View Quoted Tweet

💬13🔄1❤️19👀6532📊11

⚡ Powered by xgo.ing

Overall, I think all these reductive views of open-source are extremely dangerous. I think there are...

elvis (@omarsar0) — Sun, 19 Jul 2026 20:46:09 GMT

Overall, I think all these reductive views of open-source are extremely dangerous. I think there are important conversations to have around its effects on society. But it demands precise and balanced views with actual facts, data points, and well-informed proposals to move forward with.

💬1🔄0❤️3👀2876📊2

⚡ Powered by xgo.ing

NEW: AI Papers of the Week Collection I just released my AI Papers of the Week collection under our...

elvis (@omarsar0) — Sun, 19 Jul 2026 20:37:38 GMT

NEW: AI Papers of the Week Collection

I just released my AI Papers of the Week collection under our Resources hub.  

Now you can easily find some of the most important AI papers in one place. 

You can also use our new AI tutor to recommend top AI papers on any topic of interest or topics you are learning about on the platform. 

Learning about AI is not enough. It's important to keep up to date. So this is why we are building all these tools and resources to help with that. 

Go try it out here: academy.dair.ai/dashboard/reso…

We will update the collection every week and add new paper collections in the coming weeks. 

We have another killer feature dropping soon to read and annotate papers, including a completely new way to study and digest AI papers. Stay tuned!

Your browser does not support the video tag.

🔗 View on Twitter

💬13🔄13❤️91👀16602📊35

⚡ Powered by xgo.ing

Exactly! This is what all these heavily politicized, self-serving AI takes that are emerging against...

elvis (@omarsar0) — Sun, 19 Jul 2026 19:38:39 GMT

Exactly! This is what all these heavily politicized, self-serving AI takes that are emerging against open-source AI completely dismiss. I concede that business dynamics change and that uncertainty is uncomfortable, but if you consider what the history of open-source has been, the benefits are clear for scientific progress and humanity.

Melanie Mitchell@MelMitchell1

In the off chance than anyone cares what I think on this topic:

(1)  the open-source software movement has been enormously beneficial to socity

(2) open-weight (& better, open-data) LLMs will be essential for understanding this technology & for it to be beneficial to society.

🔗 View Quoted Tweet

💬11🔄5❤️31👀11989📊11

⚡ Powered by xgo.ing

Man, it's baffling to me that people either forget too quickly about the impact of open source or ar...

elvis (@omarsar0) — Sun, 19 Jul 2026 19:12:16 GMT

Man, it's baffling to me that people either forget too quickly about the impact of open source or are just too lazy to pick up a book and read about it. Having worked on open-source tools all these years, it's really sad to see such ignorance. Thanks for reminding people, Yann.

Yann LeCun@ylecun

@hosseeb@deanwball Soooo, releasing Linux was dumping? 
Apache, MySQL, PHP? 
HTTP, TCP/IP, OpenSSL, OpenSSH? 
Libjpeg, VLC?
The open source software stack of the mobile communication network?
Signal?
PyTorch?
Llama?

🔗 View Quoted Tweet

💬9🔄2❤️57👀16546📊12

⚡ Powered by xgo.ing

Excited for the Qwen 3.8 open-weight release. I have had real success with Qwen 3.7 for subagent wor...

elvis (@omarsar0) — Sun, 19 Jul 2026 17:52:09 GMT

Excited for the Qwen 3.8 open-weight release. I have had real success with Qwen 3.7 for subagent workflows in harnesses like Hermes Agent. They are doing something right around agentic capabilities. If Qwen 3.8 is a significant improvement, you don't want to ignore it.

Qwen@Alibaba_Qwen

Qwen3.8 is launching and going open-weight soon!🌐

With a massive 2.4T parameters, this model is continuously evolving. We believe it’s one of the most powerful model available today, compatible to leading frontier AI models , second only to Fable 5.

You don't have to wait to test it. Just now, the Qwen3.8-Max-Preview made its debut on Alibaba’s Token Plan, Qoder, and QoderWork. Be among the very first to try it out.

Can't wait to hear what you build. Stay tuned! 🚀 

Token Plan 
internationaqwencloud.com/pricing/token-…Bv
Chinplatform.qianwenai.com/pricing/token-…Rp

🔗 View Quoted Tweet

💬12🔄0❤️43👀9320📊12

⚡ Powered by xgo.ing

Open source and competition are such beautiful things. Why would anyone want to kill those things? T...

elvis (@omarsar0) — Sun, 19 Jul 2026 17:29:53 GMT

Open source and competition are such beautiful things. Why would anyone want to kill those things? The answer is clear. Let's continue to accelerate open source AI. We win when this happens.

Guillermo Rauch@rauchg

Based on internal evals:

▪️ Kimi K3 is top-tier at cybersecurity
There is chatter on X that Moonshot benchmark-overfit. These are stealth evals. Model has raw IQ.

▪️ Sol is a leap ahead in cyber capability
At a significantly higher cost, but quite remarkable still.

▪️ Fable refuses everything
We couldn’t get it to complete the run at all. What’s interesting is that Sol in comparison was much more open to helping with defensive cyber hardening

TL;DR: frontier, open-weight cybersecurity capability is here. Try it on deepsec.sh for defensive purposes.

🔗 View Quoted Tweet

💬16🔄2❤️42👀11674📊15

⚡ Powered by xgo.ing

"Open-weight models are inherently decelerationist." Among the most absurd things I have read on X ...

elvis (@omarsar0) — Sun, 19 Jul 2026 15:10:42 GMT

"Open-weight models are inherently decelerationist."

Among the most absurd things I have read on X recently. 

"Only we should build AI, and you shouldn't" is a dangerous movement and a huge scam we should be calling out. 

@TimSweeneyEpic describes it best in his reply: "Picture an executive a taco company saying this sort of thing about a new brand of tacos coming onto the market, speculating about the geopolitical and societal disruptions they anticipate as a result of advances in tacos."

I am also convinced these people don't actually use open models seriously or even try to understand their value. If you care about AI sovereignty, you would know how bad the argument is. I have never heard about this person before, but I was surprised to see such a statement given where they work. The reaction to the original post says it all.

Dean W. Ball@deanwball

Some observations on Kimi: 

It's a very good model! I don't think its performance can be explained away by distillation or anything like that. In agentic coding sessions, it seems pretty much on par with the best public models of Q1 2026. In my fairly limited use, it also seemed very token hungry. It's not obvious to me that this model is actually that cheap to run. 

I am personally surprised the Chinese state continues to allow the open sourcing of models this good, given potential risks. To be clear, I *myself* might be fine with models presenting this level of marginal risk being open weight, but I am surprised that China is fine with it. I suspect the reason they are is 75% explained by strategic blindness/lack of AGI-pilledness (the CCP is very Yann Lecun-y in its views of AI). The other 25% or so is their lack of compute for customer inference (making China's open-weight strategy an unintended byproduct of US export controls) and the normal Chinese strategy of aggressive exports. For the companies, as opposed to the government, the decision to open source is partially ideological and partially because they are behind, and they know that very few people would pay for sub-frontier models from China. 

Open-weight models are inherently decelerationist, and I'm continually surprised to see the so-called "accelerationists" so excited about open-weight models. I suspect the reason they are is that they know open-weight models are effectively ungovernable, and they simply like the overall cloak of ungovernability open-weight models create over the whole of AI. It's not a bad strategy; it reminds me of James Scott's recounting of the hill people in "the art of not being governed." Still, in the end, open-weight models deter further AI capex. 

One probable outcome of an open-weight-model-dominant world is full AI communism, which is precisely what China proposes: rather than a market product, AI is a "public good" which will ultimately be provided by the state as a kind of "digital public infrastructure." This future strikes me as a dystopian hellscape, but I've never met an open-weight models advocate who doesn't ultimately concede this is where things end. You'd be surprised how many 'accelerationists' lobbied me, while I was in government, to support an eleven or twelve-figure federally funded data center so that startups could train models at a subsidy and then give them away for free. There was no other way for AI to progress, they said. Perhaps this is the logical end state of things. Nonetheless, I find myself surprised to see supposed accelerationists excited about such an outcome. I think many of them just don't know what they're doing. Many accelerationists do not view the creation and serving of frontier models as a legitimate business. 

I would guess that the Trump Administration will at some point realize that their best strategy here would be to create large amounts of regulatory risk around the use of open-weight Chinese models. You don't need to "ban open source" (one of the dumber motifs of AI policy discussion). You just need to direct every agency to issue soft law that creates FUD. "A Federal Reserve Advisory Bulletin found that there may be backdoors in Chinese AI models." It needn't be that well justified. You just create enough regulatory risk that every regulated enterprise backs off. You probably don't want to create so much regulatory risk that you scare off the hyperscalers from serving Chinese models; this will just drive startups to sketchier providers. There's a happy middle ground here. I'd assume they will do some version of this. 

It's probably true that open-weight models of this capability make the world a bit more dangerous, but not so much more that you'll really notice. At some point the models will be capable enough that you will notice. "A nonliving, invisible, dangerous, and infinitely self-replicating agent escaped from a Chinese lab," you say? Color me shocked.

🔗 View Quoted Tweet

💬15🔄1❤️71👀11220📊17

⚡ Powered by xgo.ing

// Agents Do Not Fail Alone // Very nice open-source eval tool to check agent reliability. Lots of...

elvis (@omarsar0) — Sat, 18 Jul 2026 21:01:02 GMT

// Agents Do Not Fail Alone //

Very nice open-source eval tool to check agent reliability.

Lots of cool ideas in there.

(bookmark it)

What's the problem?

Context engineering has become central to building reliable agents but it remains almost entirely unmeasured.

Instructions, tools, memory, retrieved knowledge, guardrails, and untrusted inputs all accumulate in the context, and when that gets weak the agent drifts, hallucinates, misuses tools, becomes vulnerable to injection, and burns tokens.

This work validates context-engineering quality as an independent leading indicator of agent reliability. The measurement lives in ProofAgent-Harness, open-source infrastructure that scores context with multi-juror consensus across seven criteria, covering role clarity, guardrail coverage, instruction consistency, tool schema quality, grounding sufficiency, injection hardening, and token efficiency.

The context score is isolated from behavioral metrics and release decisions, so the prediction is not smuggled in. Holding frontier LLM agents fixed and varying only their operating context, each criterion predicts its matching outcome.

Why you want to consider doing these checks:

> Grounding sufficiency predicts hallucination resistance.
> Guardrail coverage predicts manipulation resistance.
> Tool-schema quality predicts tool use.

Paper: arxiv.org/abs/2607.14275

Learn to build effective AI agents in our academy: academy.dair.ai

💬45🔄24❤️197👀17595📊97

⚡ Powered by xgo.ing

As promised, here it is: https://t.co/0rnxKCfLil

elvis (@omarsar0) — Sat, 18 Jul 2026 20:11:46 GMT

As promised, here it is: x.com/omarsar0/statu…

elvis@omarsar0

Sharing my new skill!

It keeps track of high-signal X accounts for top AI news, papers, projects, etc.

Total gamechanger for me.

Built with X MCP tools. Give your agent the skill and tell it to generate the artifact with top stories. 

Works for Codex, Claude, Hermes, OpenClaw, or whatever you use. 

3 steps:

1. Set up X MCP - X API: docs.x.com/tools/mcp#x-mc…

2. Install skill here: github.com/dair-ai/dair-a…

3. Run prompt: "Use the x-agent-intelligence skill to build a self-contained local feed from my X MCP connection; ask for my source handles if needed, save feed.html, and validate it."

It should generate a nice, beautiful HTML artifact like the one shown in the clip. You can tune it however you want.

You can then set a schedule/automation to do this daily or whatever cadence you prefer. I have it every 4 hours. 

You will need to curate the X accounts yourself, but I have shared a few good ones under the assets. 

You can ask your agent to tune it to however you like. 

I have also shared my personal feed with our community here: academy.dair.ai/dashboard/reso…

I understand if it gets tricky to set up. Please reach out to me in the community forum. I plan to do a little tutorial or live session soon to help others reproduce the process.

You can also store the feed as a wiki, as I have in my own implementation, but that's optional. 

If you encounter any issues or have ideas on how to improve it, please open a PR.

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬1🔄0❤️3👀3482📊2

⚡ Powered by xgo.ing

Sharing my new skill! It keeps track of high-signal X accounts for top AI news, papers, projects, e...

elvis (@omarsar0) — Sat, 18 Jul 2026 20:10:59 GMT

Sharing my new skill!

It keeps track of high-signal X accounts for top AI news, papers, projects, etc.

Total gamechanger for me.

Built with X MCP tools. Give your agent the skill and tell it to generate the artifact with top stories. 

Works for Codex, Claude, Hermes, OpenClaw, or whatever you use. 

3 steps:

1. Set up X MCP - X API: docs.x.com/tools/mcp#x-mc…

2. Install skill here: github.com/dair-ai/dair-a…

3. Run prompt: "Use the x-agent-intelligence skill to build a self-contained local feed from my X MCP connection; ask for my source handles if needed, save feed.html, and validate it."

It should generate a nice, beautiful HTML artifact like the one shown in the clip. You can tune it however you want.

You can then set a schedule/automation to do this daily or whatever cadence you prefer. I have it every 4 hours. 

You will need to curate the X accounts yourself, but I have shared a few good ones under the assets. 

You can ask your agent to tune it to however you like. 

I have also shared my personal feed with our community here: academy.dair.ai/dashboard/reso…

I understand if it gets tricky to set up. Please reach out to me in the community forum. I plan to do a little tutorial or live session soon to help others reproduce the process.

You can also store the feed as a wiki, as I have in my own implementation, but that's optional. 

If you encounter any issues or have ideas on how to improve it, please open a PR.

Your browser does not support the video tag.

🔗 View on Twitter

💬45🔄73❤️773👀69273📊379

⚡ Powered by xgo.ing

I meant included. The model is still possible to use but with usage credits.

elvis (@omarsar0) — Sat, 18 Jul 2026 03:13:59 GMT

I meant included. The model is still possible to use but with usage credits.

💬1🔄0❤️2👀2014📊1

⚡ Powered by xgo.ing

Also, it sucks that it’s not available in the Pro plan. But to be honest, this model drains usage so...

elvis (@omarsar0) — Sat, 18 Jul 2026 03:12:23 GMT

Also, it sucks that it’s not available in the Pro plan. But to be honest, this model drains usage so fast that even with a Max plan I find it quite constraining.

💬4🔄0❤️6👀2970📊4

⚡ Powered by xgo.ing

I am glad they heard us. 🙏🏻

elvis (@omarsar0) — Sat, 18 Jul 2026 02:59:47 GMT

I am glad they heard us. 🙏🏻

elvis@omarsar0

I don't think Anthropic realizes how disruptive these changes are to users. I appreciate the extension, but please stop playing games. Either keep it under the subscriptions or put it under the API already.

🔗 View Quoted Tweet

💬1🔄0❤️6👀1762📊1

elvis(@omarsar0)

The hard part of multi-agent systems is getting agents to stay quiet. Put five agents on one task, ...

Peter knows something, of course. But all these loops and graphs are just terms that will consolidat...

It's clear to me that a lot of these ideas are very early. But having built my own agent orchestrato...

Dynamic workflows are a generalization of harnesses, automations, loops, routing, and graphs. It's...

Great paper on self-improving agent harnesses. (bookmark it) If you maintain a production agent ha...

Just implemented a new interface for our @dair_ai community. I first brainstormed a concept design...

In addition, something I have been building in my orchestrator system is to dynamically route to sev...

Recommended reading. Cursor Router routes tasks to the right model. Is anyone building this as ope...

I will be doing a live session on this for our community. Join us here: https://t.co/aG0VzVZyhP

It sucks that with most coding agents you have to call out the skill every single time. The agent ...

Karpathy's point about long voice rambles goes further when you add more modalities. My favorite ...

https://t.co/juAGYwoKIE

Are structured outputs in agents always good? This paper suggests that you might have to take a clo...

It looks like it only works for Claude Cowork. That's really annoying. I think it's really worth it ...

Recommended reading. "Gigatoken can tokenize the entire internet in <7 hours on one machine." ...

I agree with what this AI paper suggests. Self-improving agents should evolve their benchmarks too....

I was a bit too early on this. But I highly recommend reading it. It essentially packages two id...

Progressive disclosure in agents doesn't scale. And its benefits seems agent harness dependent. (b...

Love it how Jack keeps pushing for open-source AI. First Goose, now Buzz (looks like a great alterna...

Open weight models galore! Laguna S 2.1 went from the start of training to launch in under 9 weeks!...

Thesean just launched Ship in beta, an endpoint that cuts your LLM bill in half. You keep the model...

Not because Andrej is saying it but I think voice is goated. And you can mix it with other modaliti...

Great to see more people finding success with multimodal prompting

This is a neat feature. I wrote an article a few weeks back about how I built this into my agent o...

This is a radical improvement in how to use AI. @magicX_ai's AI Autocomplete predicts the actions a...

Octen's web search latency is pretty insane. Deep research tools like OpenAI, Gemini, Grok, and Pe...

Don’t sleep on these orchestration models. Sakana just announced Fugu-Cyber which achieves SoTA on r...

Huge launch from @tryramp. Different steps in an agent workflow can use different models. This can ...

Recommended reading. (bookmark it) Pay attention to the prices and what a combination of models c...

Something really neat about this work that might not get picked up easily: for post-training, you ma...

Wanted to add that I care about this problem. As an independent researcher, I've been working hard a...

Highly recommended. I've often claimed there's huge alpha in building agent harnesses. Turns out ...

// Global Workspace in LLMs // arXiv paper for the popular J-space work from Anthropic. (bookmark ...

A lot of young researchers will look at this and feel lost and hopeless. Some might be thinking: “I...

This is not something that anyone can just one shot with an LLM. In my experience with LLMs, it does...

That being said, I think it’s important to understand the methodology and the LLM traces behind such...

What!? It looks correct. I am convinced we are not pushing these LLMs enough. Benchmarks are simpl...

Overall, I think all these reductive views of open-source are extremely dangerous. I think there are...

NEW: AI Papers of the Week Collection I just released my AI Papers of the Week collection under our...

Exactly! This is what all these heavily politicized, self-serving AI takes that are emerging against...

Man, it's baffling to me that people either forget too quickly about the impact of open source or ar...

Excited for the Qwen 3.8 open-weight release. I have had real success with Qwen 3.7 for subagent wor...

Open source and competition are such beautiful things. Why would anyone want to kill those things? T...

"Open-weight models are inherently decelerationist." Among the most absurd things I have read on X ...

// Agents Do Not Fail Alone // Very nice open-source eval tool to check agent reliability. Lots of...

As promised, here it is: https://t.co/0rnxKCfLil

Sharing my new skill! It keeps track of high-signal X accounts for top AI news, papers, projects, e...

I meant included. The model is still possible to use but with usage credits.

Also, it sucks that it’s not available in the Pro plan. But to be honest, this model drains usage so...

I am glad they heard us. 🙏🏻