Jeremy Howard retweeted
"The models that are actually causing issues right now are all closed weight American models." -- @mitsuhiko lucumr.pocoo.org/2026/9/12/pdoo…
2
13
72
9,133
Jeremy Howard retweeted
This is the type of post I might regret later, but I really had an urge to rant about the current discourse around pacing the frontier. lucumr.pocoo.org/2026/9/12/pdoo…
42
71
654
74,461
Jeremy Howard retweeted
It's now clear that alignment is critical and won't be solved behind the closed doors of a handful of frontier labs.
So today we're launching the Open Alignment Initiative, led by @Thom_Wolf @huggingface and asking to be part of the "embedded evaluators" program that @DarioAmodei just committed to.
Let's make AI safer by making it more transparent!
So today we're launching the Open Alignment Initiative, led by @Thom_Wolf @huggingface and asking to be part of the "embedded evaluators" program that @DarioAmodei just committed to.
Let's make AI safer by making it more transparent!
We Must Pace the Frontier: I’ve written a new essay on why the AI industry should slow down, with a three-part plan for doing so.
Anthropic is unilaterally committing to the first of these steps. We’ll provide third-party evaluators with permanent, employee-level access to our systems, so that they can verify adherence to our safety measures, report on incidents, and assess models’ alignment during training.
You can read the full post here: darioamodei.com/post/we-must-p…
289
417
3,854
444,591
Jeremy Howard retweeted
Today we're launching Base Labs, a research lab by Baseten. Our mandate is to make open-source as useful as possible, and our one rule is that we publish without exception, including what fails.
Up until pretty recently I thought the way to get the world onto open models was to train them for one company at a time. @mudithj, @maxkirkby and I cofounded @parsedlabs on that bet, @baseten acquired us, and we spent the last year running their training team doing it for customers one by one. Every one of those engagements taught us something new about how models learn, forget, specialise and get cheaper, and almost none of it got spoken about. Sadly, in general that's the field's default in that the people who know the most about training have the least freedom to say it.
To be clear I don't think the closed labs are the villains here. They get to new capabilities first, which buys the rest of us time to harden the world before that stuff is everywhere, and they're the ones paying to find out what's actually possible. But I'm fairly convinced the only real advantage they have is data and scale, and their incentives point squarely at the frontier. You can't do slow, public science on how these things learn when your job is the best model by end of quarter. Someone without that pressure has to, and there are very few of those someones around. Hence Base Labs.
We have the broad remit of making open-source as useful as possible and our one rule is that we publish without exception, including what fails. The first problem is continual learning, which I have come to think is several problems wearing one name. We are also working on the open RL environments and data that open models need and currently can't get, because we have to aggregate data with the same ferocity everyone's been talking about aggregating compute. Plus a bunch of other stuff I'm genuinely excited about, eg a safety stack people can run on top of open deployments, and performance research so these things are cheap for everyone to serve.
I still think open and closed coexist, and that's the good world. It's just that coexistence isn't free, someone has to actually do the work, and this is basically what keeps me up at night. We're hiring researchers, engineers and fellows. Come help distribute the mandate of heaven!
Up until pretty recently I thought the way to get the world onto open models was to train them for one company at a time. @mudithj, @maxkirkby and I cofounded @parsedlabs on that bet, @baseten acquired us, and we spent the last year running their training team doing it for customers one by one. Every one of those engagements taught us something new about how models learn, forget, specialise and get cheaper, and almost none of it got spoken about. Sadly, in general that's the field's default in that the people who know the most about training have the least freedom to say it.
To be clear I don't think the closed labs are the villains here. They get to new capabilities first, which buys the rest of us time to harden the world before that stuff is everywhere, and they're the ones paying to find out what's actually possible. But I'm fairly convinced the only real advantage they have is data and scale, and their incentives point squarely at the frontier. You can't do slow, public science on how these things learn when your job is the best model by end of quarter. Someone without that pressure has to, and there are very few of those someones around. Hence Base Labs.
We have the broad remit of making open-source as useful as possible and our one rule is that we publish without exception, including what fails. The first problem is continual learning, which I have come to think is several problems wearing one name. We are also working on the open RL environments and data that open models need and currently can't get, because we have to aggregate data with the same ferocity everyone's been talking about aggregating compute. Plus a bunch of other stuff I'm genuinely excited about, eg a safety stack people can run on top of open deployments, and performance research so these things are cheap for everyone to serve.
I still think open and closed coexist, and that's the good world. It's just that coexistence isn't free, someone has to actually do the work, and this is basically what keeps me up at night. We're hiring researchers, engineers and fellows. Come help distribute the mandate of heaven!
Today we're announcing Base Labs, a dedicated research organization focused on advancing open-source AI.
We believe in a healthy, open frontier model ecosystem. To enable this, we are working on:
- Blue-sky research on continual learning, the science of RL, and how models learn ycle, with every experiment and recipe shared openly.
- The BaseHub Data Foundry: the highest-quality open RL environments, training data, and real-world benchmarks, built for anyone to train and benchmark on.
- Post-post training: taking open-source models and making them better, safer, and more aligned through continual post-training, built on our research, and deployed with our frontier safety stack so organizations can use open-source models with confidence.
- Making models cheaper and more performant through our model performance research.
This is a mission-driven research effort, not a commercial product. We believe the health of the open-source AI ecosystem matters and that the best way to advance it is to do science in the open.
We’re hiring engineers, researchers, and research fellows to advance this mission.
labs.baseten.co
67
64
935
129,118
Jeremy Howard retweeted
I'm in San Francisco this week. So is the Deputy Prime Minister, but we’re here for very different reasons.
He's meeting with Anthropic and OpenAI, working on Australia’s data centre pipeline.
I'm here as the co-founder of @_Firmable and a Partner at @GlitchCapital, meeting with Australian founders who moved here this year. The CGT changes came up in almost every conversation as part of the reason they left Australia to build their business in the US.
What a mess.
The AFR published two opinion pieces todaythat demonstrate this shambles.
In the first, Treasurer Jim Chalmers set out how Australia can benefit from AI if we get it right. And that’s a big if. The headline number he was spruiking was $150 billion of data centre investment by 2030.
Jim, data centres are the bottom of the stack. Concrete, power and imported GPUs. Enormous capex, very few jobs, and negligible compounding growth for our country. If that's only our AI strategy, we’re in trouble, as we're importing expensive chips, hosting someone else's product, and then buying the intelligence built on top of it at full retail.
The bigger opportunity is in using and complementing the frontier models. Products built on top of frontier models that solve real business problems. That's where jobs are created and where the productivity Chalmers is chasing actually comes from. I’m seeing it at what we’re building at Firmable, the companies we invest in at Glitch and the founders I’m meeting in the Bay Area.
In the second op-ed, @Shaun_Cartoon laid out what Chalmers is actually doing to the people who build that layer. Effective rates for founder and employee equity are doubling to 47%, which is why founders are leaving.
Australia is courting the frontier labs on one hand, and taxing the hell out of the Australian founders and builders.
You can't be a value-add exporter of AI when the value-adders are in San Francisco.
If Australia is serious about backing innovation, restore the 50% CGT discount for all businesses. This alone will generate far greater productivity gains for our country than data centres will.
He's meeting with Anthropic and OpenAI, working on Australia’s data centre pipeline.
I'm here as the co-founder of @_Firmable and a Partner at @GlitchCapital, meeting with Australian founders who moved here this year. The CGT changes came up in almost every conversation as part of the reason they left Australia to build their business in the US.
What a mess.
The AFR published two opinion pieces todaythat demonstrate this shambles.
In the first, Treasurer Jim Chalmers set out how Australia can benefit from AI if we get it right. And that’s a big if. The headline number he was spruiking was $150 billion of data centre investment by 2030.
Jim, data centres are the bottom of the stack. Concrete, power and imported GPUs. Enormous capex, very few jobs, and negligible compounding growth for our country. If that's only our AI strategy, we’re in trouble, as we're importing expensive chips, hosting someone else's product, and then buying the intelligence built on top of it at full retail.
The bigger opportunity is in using and complementing the frontier models. Products built on top of frontier models that solve real business problems. That's where jobs are created and where the productivity Chalmers is chasing actually comes from. I’m seeing it at what we’re building at Firmable, the companies we invest in at Glitch and the founders I’m meeting in the Bay Area.
In the second op-ed, @Shaun_Cartoon laid out what Chalmers is actually doing to the people who build that layer. Effective rates for founder and employee equity are doubling to 47%, which is why founders are leaving.
Australia is courting the frontier labs on one hand, and taxing the hell out of the Australian founders and builders.
You can't be a value-add exporter of AI when the value-adders are in San Francisco.
If Australia is serious about backing innovation, restore the 50% CGT discount for all businesses. This alone will generate far greater productivity gains for our country than data centres will.
25
78
373
28,179
Jeremy Howard retweeted
Papers on recurrent transformers: arxiv.org/pdf/2606.21562 (ours): train full Tr, distill -> rec Tr
arxiv.org/pdf/2608.02870 train full and rec jointly, max. coherence
arxiv.org/pdf/2606.06479 train rec Tr w predict. obj.
arxiv.org/pdf/2608.08888 train full Tr, add rec latents
arxiv.org/pdf/2608.02870 train full and rec jointly, max. coherence
arxiv.org/pdf/2606.06479 train rec Tr w predict. obj.
arxiv.org/pdf/2608.08888 train full Tr, add rec latents
2
47
317
24,813
Jeremy Howard retweeted
Claude Fable 5.1 tops the Artificial Analysis Intelligence Index but costs 20% more per task than Fable 5 despite a 75% cache read price cut
We supported @AnthropicAI with pre-release evaluation of Claude Fable 5.1. At max effort it scores 66 on the Artificial Analysis highest score we have measured, ahead of Claude Opus 5 (max, 63), Claude Fable 5 (max, 62), GPT-5.6 Sol (max, 61) and Grok 4.6 (high, 61). We evaluated the model with Anthropic's ‘default’ server-side fallback, which routes safety-flagged requests to Claude Opus 4.8 or Claude Opus 5; fallback served ~4% of output tokens across the Intelligence Index.
Key takeaways
➤ Frontier Intelligence with improvements across benchmarks: Fable 5.1 gains +4 points on the Intelligence Index over Fable 5. On HLE, Fable 5.1 scores 59.1%, ahead of the previous best of 55.5% from Claude Fable 5. It posts the narrowly highest scores we’ve seen on Terminal-Bench v2.1 (91.4%) and SciCode (62.0%), and on τ³-Banking it gains 9 points over Fable 5
➤ 75% cache read price cut, but Fable 5.1 still costs more per task: Anthropic has cut the cache read price from $1 to $0.25 per 1M cached input tokens, with standard pricing unchanged at $10/$50 per 1M input/output tokens. Fable 5.1 (max) costs $3.76 per Intelligence Index task, 20% more than Fable 5 (max), because it uses ~1.7x the output tokens. The cache cut saves ~$1.40 per task, concentrated in the agentic evaluations where the majority of input tokens are cache reads. At xhigh effort Fable 5.1 scores 65 at $2.72 per task, $1.04 less than max, but still above Claude Opus 5 (max, 63) at $2.34
➤ Claude Fable 5.1 holds the upper end of the Intelligence vs Output Tokens per Task Pareto frontier: every model variant scoring higher than GPT-5.6 Sol (medium) on the Intelligence Index is matched or beaten by a Fable 5.1 effort level on both intelligence and token usage
➤ Highest scores on agentic work tasks, but effectively tied with Opus 5: Fable 5.1 sets the highest scores we have measured on GDPval-AA v2 (1,853 Elo, +130 over Fable 5) and AA-Briefcase (1,694 Elo, +122 over Fable 5), our agentic knowledge work evaluations. Against Claude Opus 5 the GDPval-AA v2 lead is within the confidence interval and AA-Briefcase (1,685) is effectively tied, with Fable 5.1 ahead on analytical quality and rubric correctness, but behind on presentation
Other model details:
➤ Context window: 1 million tokens, supporting image and text inputs as with Anthropic’s other recent launches
➤ Pricing: Fable 5.1 retains the $10/$50/$12.5 input, output, and cache write prices per million tokens from Fable 5, but cache hits have been reduced to $0.25 per million tokens, a 75% relative reduction from before that will materially reduce agentic workload costs
We supported @AnthropicAI with pre-release evaluation of Claude Fable 5.1. At max effort it scores 66 on the Artificial Analysis highest score we have measured, ahead of Claude Opus 5 (max, 63), Claude Fable 5 (max, 62), GPT-5.6 Sol (max, 61) and Grok 4.6 (high, 61). We evaluated the model with Anthropic's ‘default’ server-side fallback, which routes safety-flagged requests to Claude Opus 4.8 or Claude Opus 5; fallback served ~4% of output tokens across the Intelligence Index.
Key takeaways
➤ Frontier Intelligence with improvements across benchmarks: Fable 5.1 gains +4 points on the Intelligence Index over Fable 5. On HLE, Fable 5.1 scores 59.1%, ahead of the previous best of 55.5% from Claude Fable 5. It posts the narrowly highest scores we’ve seen on Terminal-Bench v2.1 (91.4%) and SciCode (62.0%), and on τ³-Banking it gains 9 points over Fable 5
➤ 75% cache read price cut, but Fable 5.1 still costs more per task: Anthropic has cut the cache read price from $1 to $0.25 per 1M cached input tokens, with standard pricing unchanged at $10/$50 per 1M input/output tokens. Fable 5.1 (max) costs $3.76 per Intelligence Index task, 20% more than Fable 5 (max), because it uses ~1.7x the output tokens. The cache cut saves ~$1.40 per task, concentrated in the agentic evaluations where the majority of input tokens are cache reads. At xhigh effort Fable 5.1 scores 65 at $2.72 per task, $1.04 less than max, but still above Claude Opus 5 (max, 63) at $2.34
➤ Claude Fable 5.1 holds the upper end of the Intelligence vs Output Tokens per Task Pareto frontier: every model variant scoring higher than GPT-5.6 Sol (medium) on the Intelligence Index is matched or beaten by a Fable 5.1 effort level on both intelligence and token usage
➤ Highest scores on agentic work tasks, but effectively tied with Opus 5: Fable 5.1 sets the highest scores we have measured on GDPval-AA v2 (1,853 Elo, +130 over Fable 5) and AA-Briefcase (1,694 Elo, +122 over Fable 5), our agentic knowledge work evaluations. Against Claude Opus 5 the GDPval-AA v2 lead is within the confidence interval and AA-Briefcase (1,685) is effectively tied, with Fable 5.1 ahead on analytical quality and rubric correctness, but behind on presentation
Other model details:
➤ Context window: 1 million tokens, supporting image and text inputs as with Anthropic’s other recent launches
➤ Pricing: Fable 5.1 retains the $10/$50/$12.5 input, output, and cache write prices per million tokens from Fable 5, but cache hits have been reduced to $0.25 per million tokens, a 75% relative reduction from before that will materially reduce agentic workload costs
110
191
2,080
507,030
Jeremy Howard retweeted
RWKV-7 G1j (100% RNN) release 🙂 much better at agent/coding/STEM and everything. Demo: rwkv.com Weights: huggingface.co/BlinkDL/rwkv7-…
RWKV-7 G1i (100% RNN) release 🙂 Demo: rwkv.com Weights: huggingface.co/BlinkDL/rwkv7-…
3
12
128
24,895
Jeremy Howard retweeted
my mom is using gemini on her android phone to write angry letters to companies/banks who did her wrong by law and they are all giving in, as gemini cites the law and i think this is beautiful.
62
132
5,507
176,221
Jeremy Howard retweeted
At 1400x cheaper, I know I'm sticking to embeddings (dense, sparse, multi-vector), plus hybrid (incl. bm25) and rerankers.
Perhaps I'd even use listwise cross-encoders, they seem interesting.
Perhaps I'd even use listwise cross-encoders, they seem interesting.
can LLMs replace embedding models? in a new work "embedder's dilemma" we find LLMs now beat embedding models -- but at much higher cost. when to choose which?...
📜 arxiv.org/abs/2608.12875
15
27
216
24,567
Jeremy Howard retweeted
How Anthropic's new results post would read without the PR:
Claude orchestrated open-source protein design models, PXDesign, RFdiffusion, Genie, BoltzGen, from a 30k-token expert prompt and 12,500 H100-hours of compute, and designed binders against 14 of 15 targets. Hit rates of 22–35% against a 10–15% baseline, where some of those tools already report similar numbers on their own.
The orchestration is genuinely impressive. But the open-source models did most of the lifting, and they came from the Baker lab, Columbia, MIT, ByteDance Seed, and most of them were already wet-lab validated before Claude touched them.
Which also sets the ceiling. All these generators share a single PDB-shaped training distribution, so calling four of them doesn't diversify away the blind spot, since they fail together. The targets that worked are the well-studied ones.
So the valid claim is that an agent can now drive this stack competently in the regime where the stack already works.
Instead, we got this announcement:
Claude orchestrated open-source protein design models, PXDesign, RFdiffusion, Genie, BoltzGen, from a 30k-token expert prompt and 12,500 H100-hours of compute, and designed binders against 14 of 15 targets. Hit rates of 22–35% against a 10–15% baseline, where some of those tools already report similar numbers on their own.
The orchestration is genuinely impressive. But the open-source models did most of the lifting, and they came from the Baker lab, Columbia, MIT, ByteDance Seed, and most of them were already wet-lab validated before Claude touched them.
Which also sets the ceiling. All these generators share a single PDB-shaped training distribution, so calling four of them doesn't diversify away the blind spot, since they fail together. The targets that worked are the well-studied ones.
So the valid claim is that an agent can now drive this stack competently in the regime where the stack already works.
Instead, we got this announcement:
Many drugs work by binding to a specific target in the body and blocking or changing what it does. An important first step in the drug development process is designing a molecule that can bind tightly to its target. Traditionally, that's meant weeks or months of expert work per a large number of candidates to identify the few that work.
We wanted to test if Claude could successfully design novel protein binders from scratch (also called de novo design). With a protein design prompt written by a human expert, Claude autonomously designed protein binders against 14 out of 15 targets.
We then worked with Adaptyv Bio and Twist Bioscience, who independently built and tested the proteins Claude designed.
74
253
2,261
320,761
Jeremy Howard retweeted
Regarding Sentence Transformers v6.0 update:
A dense model compresses a whole text into one vector, then compares two vectors.
A multi-vector model keeps one vector per token, scores every query token against every document token, takes the best match for each, and sums those.
A dense model compresses a whole text into one vector, then compares two vectors.
A multi-vector model keeps one vector per token, scores every query token against every document token, takes the best match for each, and sums those.
🚨I've just released Sentence Transformers v6.0!
MultiVectorEncoder joins the family: ColBERT-style late interaction models are now a first-class model type, for training, inference & interpretation, alongside dense, sparse & reranker models.
Big thread 🧵
7
15
88
12,894
Jeremy Howard retweeted
LLMs increase the complexity of codebases. They duplicate methods, write overdefensive code against impossible edge cases, & overoptimize too early. Can further training fix this? Naur's “Programming as Theory Building” says no. -- @pol_avec
answer.ai/posts/2026-08-… 1/
answer.ai/posts/2026-08-… 1/
13
53
393
20,491
Jeremy Howard retweeted
I am a big fan of both Andy Hall and Alan Rozenshtein work and public commentary on AI. I am sure that they will continue to do important work as part of this new team at Anthropic, and sincerely congratulate them on what I am sure is an exciting next step for them personally.
At the same time, I think both of them going into Anthropic is emblematic of a broader trend of the frontier AI companies grabbing up a truly remarkable number of previously independent voices and thinkers in AI.
When massive events happen in AI - e.g. the DOW <> Ant conflict - the public, media, and elected officials will be looking for independent commentary and analysis from thinkers who are not at the frontier labs. The number of folks they can go to for that sort of independent commentary seems like it is going down precipitously.
This seems like a real problem, and worth thinking more systemically about how we can create sufficiently attractive positions outside of companies - e.g. at independent nonprofits, in government - to attract the sorts of folks like Andy and Alan who are going into Anthropic.
Anyway, each of these decisions individually is reasonable and not worth overly fixating on, and I do believe that both of them will do good work as part of this team (and have access to novel information that will help them do research) but zooming out I am worried that we are losing independent voices at the exact time we need more of them not fewer.
At the same time, I think both of them going into Anthropic is emblematic of a broader trend of the frontier AI companies grabbing up a truly remarkable number of previously independent voices and thinkers in AI.
When massive events happen in AI - e.g. the DOW <> Ant conflict - the public, media, and elected officials will be looking for independent commentary and analysis from thinkers who are not at the frontier labs. The number of folks they can go to for that sort of independent commentary seems like it is going down precipitously.
This seems like a real problem, and worth thinking more systemically about how we can create sufficiently attractive positions outside of companies - e.g. at independent nonprofits, in government - to attract the sorts of folks like Andy and Alan who are going into Anthropic.
Anyway, each of these decisions individually is reasonable and not worth overly fixating on, and I do believe that both of them will do good work as part of this team (and have access to novel information that will help them do research) but zooming out I am worried that we are losing independent voices at the exact time we need more of them not fewer.
Wonderful to have Andy Hall (@ahall_research) joining as a founding member of our new team at @AnthropicAI, working on AI and the rule of law. Few people have done such original and illuminating work on the impact of AI on political life. If you don't already read his Free Systems Substack, you should start. Together with Alan Rozenshtein (@ARozenshtein) and Justin Curl (@curl_justin), that's a dream team.
13
28
238
45,096
The amazingly fast (only 30M!) and extremely accurate @answerdotai ColBERT model is now supported in Sentence Transformers to make it easy to create and query embedding indexes locally directly in Python! 🥳
🚨I've just released Sentence Transformers v6.0!
MultiVectorEncoder joins the family: ColBERT-style late interaction models are now a first-class model type, for training, inference & interpretation, alongside dense, sparse & reranker models.
Big thread 🧵
33
58
530
61,918
@answerdotai h/t to @bclavie for his amazing work on this model.
7
9
111
14,238
Jeremy Howard retweeted
Similar story. My company Support.com was one of the handful of cos that could go public in the end of the dotcom IPO window bc we had $170m in bookings. Accel went on a jihad to replace me as CEO bc i was 32 and had never run a public co. They literally couldnt find a single reason other than my lack of age and experience and i had the junior partner who needed to prove he had the mojo to change out one of ‘his’ ceo’s. I regret that i eventually caved since it led to such a lower $$ outcome with the grey haired HP ceo who had no ability to navigate the multiple reinventions needed. Thats why my advice to founders is 1) keep control at all costs and 2) take passive investors (like yuri milner/dst) over the VC builders who ultimately want to put their own stamp on your co.
Travis Kalanick reveals Bill Gurley pressured him to take a $6 billion valuation. Two months later, he raised at $17.5 billion.
"Gurley was convinced it was the end of the world and we've got to raise, and just take your first term sheet, just f*cking take it."
"We got to a dn't talk to him. If he reached out to me, I would talk to him, but I just wouldn't try to engage because it was always that. So Emil Michael would handle this part. He would handle a lot of the investor relations."
"Anyways, two months later, we got a round done at a $17.5 billion pre."
"We were already planning IPO stuff. [Benchmark Capital] didn't know it. But they also never expressed the desire to be like, 'Hey, dude, you gotta IPO.' It was never brought up."
"If you think the world's about to end and you've got this thing that's worth a ton of money, you could get to a weird place. Especially when you're a VC, you don't have any control. You're not running anything. So now your entire outcome in life is due to this guy."
"You could get into a place where you're running a war room to destroy him. And that's kind of what happened."
49
82
1,838
341,908
After leaving AI to do a masters in immunology and start a 2nd PhD (in microbiology) I'm beyond thrilled that @math_rachel is returning to her original AI home! 🥳
She founded the Center for Applied Data Ethics *years* before most people realised AI would transform everything.
She founded the Center for Applied Data Ethics *years* before most people realised AI would transform everything.
As the backlash against AI was growing stronger (and as my friends were becoming more fervently anti-AI), I decided to return to an increasingly hated field. Why? I agree there is a lot that is terrible about AI. 1/
fast.ai/posts/2026-08-…
20
31
674
94,074
