We often talk about AI having the potential for real positive impact on peoples lives. But, we need to bring AI into the lab doing the actual work, not just reading papers (...like faculty?).
Embodied AI in the lab, is the next frontier and this is the team to make it happen.
Embodied AI in the lab, is the next frontier and this is the team to make it happen.
💥 Introducing Warp Sciences, a new company applying AI and robotics to help the world develop new medicines faster.
We’ve raised a $215M Series A led by @lightspeedvp to bring frontier AI out of the browser and into the lab, and build scientific capacity here in America.
I co-founded Warp with my long-time friend @dpatil, FAIR co-founder @manohar_paluri, and drug hunter Sol Reisberg. The early team is full of incredible people like @bkspears9, who helped achieve Q>1 fusion at LLNL.
1
2
15
1,265
Most memory benchmarks test recall and not synthesis.
My former students just launched an exciting new memory benchmark that evaluates whether agents can continually learn from past experiences.
My former students just launched an exciting new memory benchmark that evaluates whether agents can continually learn from past experiences.
Introducing FrontierMemory: a new way to evaluate agent memory through downstream task performance.
While existing memory benchmarks focus on recall (saturated via agentic search), FrontierMemory focuses on how agents learn from experience.
3
3
29
1,621
Joey Gonzalez retweeted
🚨🚨🚨 vLLM is all you need to reproduce DeepSeek’s performance.
With Wide-EP and EPLB, vLLM can match DeepInfer’s decode performance within 6% and generate $39/gpu/hr in revenue at DeepSeek’s off-peak pricing, at 200 P90 TPS on the AgentX workload.
Specifically:
• Under guration (DeepSeek-V4.1-Flash, 131K context, nextn=3), vLLM on GB200 achieves throughput within 6% of DeepInfer’s published results across batch sizes 8–256.
• On AgentX, vLLM pushes the Pareto frontier further on GB300, reaching up to 1M tokens/s/GPU and generating $38.96/gpu/hr in revenue at 196 P90 TPS, based on DeepSeek’s off-peak API pricing.
• Also, vLLM extended low-latency kernel support for DeepSeek-V4.1-Flash, achieving up to 826 TPS per user on AgentX.
With Wide-EP and EPLB, vLLM can match DeepInfer’s decode performance within 6% and generate $39/gpu/hr in revenue at DeepSeek’s off-peak pricing, at 200 P90 TPS on the AgentX workload.
Specifically:
• Under guration (DeepSeek-V4.1-Flash, 131K context, nextn=3), vLLM on GB200 achieves throughput within 6% of DeepInfer’s published results across batch sizes 8–256.
• On AgentX, vLLM pushes the Pareto frontier further on GB300, reaching up to 1M tokens/s/GPU and generating $38.96/gpu/hr in revenue at 196 P90 TPS, based on DeepSeek’s off-peak API pricing.
• Also, vLLM extended low-latency kernel support for DeepSeek-V4.1-Flash, achieving up to 826 TPS per user on AgentX.
38
98
1,040
111,931
One of my students just released some impressively fast kernels for modern variants of linear attention on TPUs.
Now we just need some TPUs to do the actual model architecture research, @GoogleDeepMind want to collaborate?
Now we just need some TPUs to do the actual model architecture research, @GoogleDeepMind want to collaborate?
Releasing FLAT: Flash Linear Attention on TPUs 🚀
Pallas training kernels for Linear Attention, Mamba-2 and GLA. On one TPU v6e, forward + backward is up to 3x faster than FLA on an H100 for GLA, and 1.5x faster for Mamba-2.
🧵
0
5
79
5,322
Joey Gonzalez retweeted
🧵 More stable PPO for LLM post-training.
Introducing EasyPPO: just fix the critic.
❌ No new actor loss
❌ No new policy algorithm
✅ Actor update unchanged
EasyPPO delivers better results with zero training collapse across all our experiments.
Introducing EasyPPO: just fix the critic.
❌ No new actor loss
❌ No new policy algorithm
✅ Actor update unchanged
EasyPPO delivers better results with zero training collapse across all our experiments.
10
43
308
42,582
Joey Gonzalez retweeted
an academic pre-training run that was spread across a bunch of different clusters guerrilla style - love it
We pretrained a 2.3B MoE (360M active) Hybrid Mamba-2 that lands within a few points of Llama-3.2-3B using <1% of its pretraining FLOPs.
No dedicated cluster. The run hopped between H100s, A100s, V100s (yes, V100s) and TPU v5p/v6e on a single codebase.
Meet Rigel 🧵
4
3
25
2,695
We just trained a 2.3B MoE Hybrid Mamba model that competes with a Llama-3.2-3B dense using less than 1% of the pre-training FLOPS.
We may not have a lot of resources in academia but we make up for it with creative students who find ways to train innovative new architectures across whatever hardware they can find.
If anyone wants to help, we would love to scale this research to the next rung of the training ladder.
We may not have a lot of resources in academia but we make up for it with creative students who find ways to train innovative new architectures across whatever hardware they can find.
If anyone wants to help, we would love to scale this research to the next rung of the training ladder.
We pretrained a 2.3B MoE (360M active) Hybrid Mamba-2 that lands within a few points of Llama-3.2-3B using <1% of its pretraining FLOPs.
No dedicated cluster. The run hopped between H100s, A100s, V100s (yes, V100s) and TPU v5p/v6e on a single codebase.
Meet Rigel 🧵
1
6
78
8,156
The agent ecosystem is evolving rapidly and we are trying to capture a picture of where we are and where things are headed.
We're now collecting responses for the 2026 survey of agents.
berkeley.qualtrics.com/jfe/form/SV_3q…
Help us create a more complete picture of how agent systems are es remain, and where future research may be most impactful.
We're now collecting responses for the 2026 survey of agents.
berkeley.qualtrics.com/jfe/form/SV_3q…
Help us create a more complete picture of how agent systems are es remain, and where future research may be most impactful.
1
10
16
3,643
Joey Gonzalez retweeted
Our social systems and institutions confuse me all the time. 😕
AI may make them even more complicated. What if it also represents us inside them?
We already delegate representation constantly - your employer negotiates your insurance, your lawyer negotiates your contract, your representatives advocate for your interests.
Wrote up some early thinking on what financial markets can teach us here, what we found in our recent paper Cooperate to Compete (now accepted to EMNLP), and why one of the hardest open problems may be designing the structures agents negotiate inside rather than the agents themselves.
Read my first research blog below!! I would love to hear any thoughts on how folks are thinking about this.
Blog: abbykoneill.github.io/posts/represen…
AI may make them even more complicated. What if it also represents us inside them?
We already delegate representation constantly - your employer negotiates your insurance, your lawyer negotiates your contract, your representatives advocate for your interests.
Wrote up some early thinking on what financial markets can teach us here, what we found in our recent paper Cooperate to Compete (now accepted to EMNLP), and why one of the hardest open problems may be designing the structures agents negotiate inside rather than the agents themselves.
Read my first research blog below!! I would love to hear any thoughts on how folks are thinking about this.
Blog: abbykoneill.github.io/posts/represen…
0
6
22
1,744
This is super cool!
@pgasawa I have been thinking about agents that move beyond turns. How would an agent know to interrupt a broken bash script as it is producing output or interrupt a user typing a confusing question? Shouldn't agents be able to think and listen continuously?
I am really excited to see @andykonwinski has already started to build such a system.
@pgasawa I have been thinking about agents that move beyond turns. How would an agent know to interrupt a broken bash script as it is producing output or interrupt a user typing a confusing question? Shouldn't agents be able to think and listen continuously?
I am really excited to see @andykonwinski has already started to build such a system.
Introducing Headlong, an open source microharness for persistent agents: self-guided agents that think continuously.
Most agent harnesses are reactive: you send a task, the agent completes it, and then it sits frozen until the next request. Cron jobs and heartbeats wake it up to it back to sleep.
A Headlong agent is never asleep. It keeps generating thoughts about whatever it decides is interesting, in a self-guided loop inspired by human inner monologue. Your message doesn't start a session. It's one more observation that lands in the agent's thought stream, and the agent decides if and when to reply.
Headlong is built on the idea of persistent agency: continuous inner thought generation between external interactions. The agent sets its own interests and priorities, comes up with its own projects, and sometimes pings you unprompted with progress.
To keep our prototype as simple and small as possible, we implemented Headlong as a microharness: a complete agent harness in under 10K lines of Bash, organized as a handful of small executables. It includes a loop that generates the next thought, shellm (a recursive language model written in Bash), a trajectory stored as a DAG of jsonl files, and context as a projection of that trajectory.
We've been running one Headlong agent internally at Laude for several weeks. The whole team talks to it over Slack and Telegram, and every conversation lands in its single stream of thought. It works in its own fork of Headlong and we've pulled over 50 of its commits into main.
One night, with nobody talking to it, it went back to check whether a recall process it had built was actually wired into its mind, found that it wasn't, diagnosed and fixed the bug, and verified the fix end to end. 48 minutes, no human asked for the fix or was in the loop at any point. Every step is a timestamped line in its log.
Things broke too, and we wrote those up. Background thinking costs us $1 to $2 an hour, our agent stopped its own service three times by accident, and self-delegation died on day one. Details in the post.
One line installs everything and starts an agent. Use a dedicated sandbox and spend-capped API key; it runs real shell commands and thinks around the clock.
Headlong is research software, be careful!
curl -fsSL headlong.ai/install.sh | bash
Launch post: laude.org/updates/headlo…
Repo: github.com/laude-institut…
Headlong is a @LaudeInstitute / MIT collaboration.
2
1
39
6,716
We need more people doing this! Great work @percyliang !
🚢 Marin 535B-A23B started training this week! As usual, the whole process is open.
Voyage plan: pretraining (80%) + midtraining (20%) on 18.75T tokens on 11 x GB200 NVL72 for ~3 months (2.7e24 FLOPs). Post-training will follow.
Before kicking off the run, we trained a 4-rung B-A61M (48B tokens) to 27.7B-A1.2B (926B tokens) to debug issues, and to make a forecast of our hero run. This is by far our biggest run, so definitely expecting the unexpected.
0
4
50
6,254
Joey Gonzalez retweeted
🚢 Marin 535B-A23B started training this week! As usual, the whole process is open.
Voyage plan: pretraining (80%) + midtraining (20%) on 18.75T tokens on 11 x GB200 NVL72 for ~3 months (2.7e24 FLOPs). Post-training will follow.
Before kicking off the run, we trained a 4-rung B-A61M (48B tokens) to 27.7B-A1.2B (926B tokens) to debug issues, and to make a forecast of our hero run. This is by far our biggest run, so definitely expecting the unexpected.
Voyage plan: pretraining (80%) + midtraining (20%) on 18.75T tokens on 11 x GB200 NVL72 for ~3 months (2.7e24 FLOPs). Post-training will follow.
Before kicking off the run, we trained a 4-rung B-A61M (48B tokens) to 27.7B-A1.2B (926B tokens) to debug issues, and to make a forecast of our hero run. This is by far our biggest run, so definitely expecting the unexpected.
106
576
4,215
990,732
Joey Gonzalez retweeted
New in @berkeley_ai blog. Cao et al. create new kernel optimizer K-search and use it to auto adapt CUDA to MLX. Their attention translation measures 0.97x speed to Apple’s native implementation. Translate your code with their OSS tool. Available now.
bair.berkeley.edu/blog/2026/07/2… 🐻📄
bair.berkeley.edu/blog/2026/07/2… 🐻📄
1
2
9
3,721
Joey Gonzalez retweeted
KV cache management is arguably one of the hardest problems in LLM inference, and advancing it requires the community to learn from and build on one another’s work. Great to see that happening. Amazing work from @lightseekorg on adopting and evolving vLLM's hybrid memory allocator!
The Jenga paper from @ChenZha62999224 is especially relevant today as attention architectures become increasingly complex.
The Jenga paper from @ChenZha62999224 is especially relevant today as attention architectures become increasingly complex.
TokenSpeed scheduler originally used a Radix Tree–based memory pool. As new attention patterns became more complex, we deprecated the Radix Tree–based design and implementation, and rebuilt the memory manager with a flat, block-based KV cache architecture.
TokenSpeed now uses a with heterogeneous views, making it easier to support different attention mechanisms. Similar approaches have also been explored in the community, including @vllm_project's Jenga and LMDeploy’s TurboMind.
This redesign happened around the release of TML’s Inkling, so we were able to support Inkling from day one with the new architecture. We are excited to keep building with the open-source community. 👇
6
49
485
42,591
Joey Gonzalez retweeted
"We're spending about five figures a month on coding agents. I think it's making the team more productive. I'm not sure. But I think."
That's from a VP of engineering running a team of ~100. It's obvious that coding agents are faster than human engineers, but it's not obvious how you manage your spend.
For decades, the constraint on shipping software was headcount. You had N engineers, you pointed them at the highest-priority work, and you shipped what you shipped.
Today you can spend 2-3x the cost to produce the same software in a tenth of the time. The recent port of Bun from Zig to Rust took 11 days and $165k in API costs. That's obviously a net improvement over what it would have cost to have engineers do it by hand, but choosing to do that requires a new calculus.
Because faster production is not the same as better economics.
The fact that you can generate more code, faster, doesn't mean you should. Every task is now a choice: which model, how much speed, how much accuracy, how much human guidance. Get those parameters wrong and you pay a lot of money to build the wrong thing very fast.
The challenge for engineering teams has always been picking the right things to build. Agents didn't fix that problem — it's faster but just as expensive: open.substack.com/pub/frontierai…
That's from a VP of engineering running a team of ~100. It's obvious that coding agents are faster than human engineers, but it's not obvious how you manage your spend.
For decades, the constraint on shipping software was headcount. You had N engineers, you pointed them at the highest-priority work, and you shipped what you shipped.
Today you can spend 2-3x the cost to produce the same software in a tenth of the time. The recent port of Bun from Zig to Rust took 11 days and $165k in API costs. That's obviously a net improvement over what it would have cost to have engineers do it by hand, but choosing to do that requires a new calculus.
Because faster production is not the same as better economics.
The fact that you can generate more code, faster, doesn't mean you should. Every task is now a choice: which model, how much speed, how much accuracy, how much human guidance. Get those parameters wrong and you pay a lot of money to build the wrong thing very fast.
The challenge for engineering teams has always been picking the right things to build. Agents didn't fix that problem — it's faster but just as expensive: open.substack.com/pub/frontierai…
0
3
4
692
Joey Gonzalez retweeted
Best Paper Award to our ICML Workshop paper 'Fantastic Adaptive Taxonomies and How to Use Them'.
Let me summarize briefly what the paper is about: Failure taxonomies are becoming increasingly important.
We show that failure taxonomies can be used in multiple ways: 1. As a for best-of-N judges, 2) as a mutation feedback mechanism in optimization loops and 3) as runtime feedback for coding agents.
Previously people used MAST and other hand-made fixed failure taxonomies. In this paper we show how to create the failure taxonomy adaptively and dynamically: We observe agent rollouts and create an adaptive failure taxonomy, bespoke to the agent weaknesses and task challenges. These adaptive taxonomies give a massive boost: On Terminal Bench 2 we get 89.9% with Opus 4.6 / Forgecode harness and the adaptive taxonomy used with a Best-of-N Judge, outperforming fixed taxonomies by 15%. (1/n)
Let me summarize briefly what the paper is about: Failure taxonomies are becoming increasingly important.
We show that failure taxonomies can be used in multiple ways: 1. As a for best-of-N judges, 2) as a mutation feedback mechanism in optimization loops and 3) as runtime feedback for coding agents.
Previously people used MAST and other hand-made fixed failure taxonomies. In this paper we show how to create the failure taxonomy adaptively and dynamically: We observe agent rollouts and create an adaptive failure taxonomy, bespoke to the agent weaknesses and task challenges. These adaptive taxonomies give a massive boost: On Terminal Bench 2 we get 89.9% with Opus 4.6 / Forgecode harness and the adaptive taxonomy used with a Best-of-N Judge, outperforming fixed taxonomies by 15%. (1/n)
8
17
126
38,922
Joey Gonzalez retweeted
Intelligence sufficient for knowledge work is rapidly becoming free. What does this mean for the future of data systems and agents? A new blog post outlines a new research agenda @Berkeley_EECS; 🧵 bair.berkeley.edu/blog/2026/07/0…
2
18
54
21,330
Joey Gonzalez retweeted
⚙️ Speaker Spotlight: Joseph Gonzalez
Every time you use an LLM — or check which model tops the leaderboard — there's a good chance his research made it possible.
@profjoeyg is a professor at @UCBerkeley and co-director of the Sky Computing Lab, whose group built some of the I systems of the LLM era.
🔹 Professor of Computer Science at @UCBerkeley · co-director, Sky Computing Lab
🔹 His group created vLLM, Chatbot Arena (now LMArena), Gorilla & Vicuna
🔹 Co-founder of RunLLM · previously co-founded Turi/GraphLab (acquired by Apple)
🔹 A leading voice on LLM agents, serving & how we evaluate AI
While others build the models, he builds the systems that run them — and the arenas that rank them.
Hear Joseph on stage at AGI Summit SF 2026.
📅 July 18–19, 2026
📍 Palace of Fine Arts, San Francisco
🎟 Tickets: agisummit.ai
🏷️ 15% off with code GenAI-26
#agisummit #aiareall #AIsystems
Every time you use an LLM — or check which model tops the leaderboard — there's a good chance his research made it possible.
@profjoeyg is a professor at @UCBerkeley and co-director of the Sky Computing Lab, whose group built some of the I systems of the LLM era.
🔹 Professor of Computer Science at @UCBerkeley · co-director, Sky Computing Lab
🔹 His group created vLLM, Chatbot Arena (now LMArena), Gorilla & Vicuna
🔹 Co-founder of RunLLM · previously co-founded Turi/GraphLab (acquired by Apple)
🔹 A leading voice on LLM agents, serving & how we evaluate AI
While others build the models, he builds the systems that run them — and the arenas that rank them.
Hear Joseph on stage at AGI Summit SF 2026.
📅 July 18–19, 2026
📍 Palace of Fine Arts, San Francisco
🎟 Tickets: agisummit.ai
🏷️ 15% off with code GenAI-26
#agisummit #aiareall #AIsystems
1
1
12
501