RL @PrimeIntellect Prev @amazon AGI SF Lab, @Apple, @SimularAI, @UCSB_AI
San Francisco, CA
Joined June 2024
- Tweets686
- Following128
- Followers2.4K
- Likes1.5K
Once again, Simular is at the top of OSWorld! Simular’s CUA research has always been SOTA, and defining the pareto frontier on the hardest benchmarks.
Incredible work by the team as always.
Our computer-use agent @sai_borg just beat Opus 5 and GPT-5.6 Sol on OSWorld 2.0.
Sai scored 73% on the CUA benchmark where each complex task takes a skilled human over an hour to execute.
And Sai did it at about 2/3 the cost per task of Opus and GPT 💸
The unlock is our neurosymbolic framework that pairs the exploratory power of neural networks with the logic of symbolic code, resulting in reliability and cost saving.
Read the full report: simular.ai/articles/sai-tops…
Frontier data is a research problem.
For example, some of the biggest academic CUA papers are solely focused on data synthesis for SFT (OpenCUA, from last August) and RL (CUA-Gym, from this June). Good data requires diligent researchers and engineers.
By next year, I'd expect data talent to look increasingly similar to the current research scientist talent. Already hear small stories of data talent poaching wars and such.
With the latest GDM news, it seems ALL the FAANGs are struggling to push out frontier models. Not to say it's impossible, but rather large companies are inherently unfit.
1 year ago, GDM was a staple in the "big 3 frontier labs". 2 years ago, the Llama efforts were open-source SOTA. Yet both fell victim to continually bleeding talent, reorgs, and layoffs.
All the Mag 7 (Microsoft, Apple, Amazon, Nvidia, Google, Meta) either struggled remaining at the frontier, or flat-out never reached frontier. The one exception is Meta who realized this and structured everyone around TBD, but that's an unprecedented investment.
The reason large companies feel "inherently unfit" is because you have thousands of people who suddenly want to touch the frontier model efforts. At the management level, orgs get bloated, scope gets messy, and compromises start piling. All the OG researchers and leaders start to rage quit.
Large companies end up with too many people wanting to own too much, without fully understanding how much time and attention to engineering detail is required to become frontier.
Feels like the most successful models come from a "small team owns the full recipe" mentality. For all the other RE/RS outside this core team, their scope ends up becoming more of a supporting/smaller role (i.e., ONLY working on a data/evals vertical, or ONLY working on compute infra). Although unfortunate for most, accepting your role ensures core researchers get the compute they need and remain undistracted from politics.
You can just hillclimb things.
Insane that all these DeepSeek-V4-Flash gains are presumably just from better post-training data.
They admit to not changing model architecture, and the Deepseek V4 report was already doing the standard learning algos (SFT + RL specialists, then MOPD into final model), so I don't imagine much changed there. Likely the main improvement is just more hard RL tasks.
🚀 DeepSeek-V4-Flash Official API is now LIVE in public beta!
🔷 We’ve massively upgraded its Agent capabilities—benchmark scores are now far surpassing the V4-Pro-Preview. Check out the massive performance leap below! 👇
🔷 The official V4-Flash now natively supports the Responses API format and is fully adapted for Codex!
Check out the configuration details in our official API docs: api-docs.deepseek.com/quick_…
👀Everyone (@Everyone) has left Amazon AGI SF Lab
Really been enjoying the paradigm shift and advocating for code-first CUA approaches. Frameworks a year ago were GUI first, code as a subagent. Now, it's the opposite: Code first, GUI as a subagent.
Code/tool calls are often just as powerful and more reliable than GUI, especially for those commonly used knowledge work apps with well supported libraries + APIs (google suite, office suite, notion, slack, etc.)
The very first OSWorld 2.0 rollout I ever looked at was GPT 5.5 spamming terminal commands. The task involved email, spreadsheet, and calendar apps, but it just opened a terminal and types away.
Feels like many people treat computer use as GUI first, and actuation via clicking and typing. However, CUA really means "whatever actuation gets the job done on a computer". Models are good at code, so it makes sense their primary interface is code.
This quoted post is unavailable.
Frontier intelligence is open: some food for thought from the K3 report.
- Feels primarily like an infra and arch paper. Likely the most important things to nail when you scale from 1T to 3T
- Seems like they focus a lot on webdev RL and use expert-curated web dev tasks. K3 beats Fable and achieve SOTA on evals like WebDev Arena and Design Arena. I've said before that distillation alone isn't enough for SOTA, you also need high fidelity RL tasks to greatly surpass your supervised boundary
- They bundle many domains into one MOPD teacher, such as throwing reasoning, search, and knowledge work all into one "general task" teacher. Probably having too many teachers makes merging difficult (for example, Nemotron 3 Ultra struggled with this and required a 'MOPD Warmup' stage)
- Instead of having one MOPD teacher for each vertical, they have an MOPD teacher for each (vertical, thinking budget) pair for better adaptive reasoning
- No RL length penalty. Instead, they use a per-task budget based on heuristics and a "human-in-the-loop" pipeline, and just penalize any rollout that exceeds the budget
- Scaled their personal assistant tasks (i.e., knowledge work/CUA data) to the order of thousands of steps
Releasing the model weights and technical report of Kimi K3.
Kimi K3 is our most capable model: a 2.8T MoE model with native visual understanding and a 1M-token context window.
New model architecture: 2.5x the intelligence per unit of compute, not just more params.
Alongside Kimi K3, we're opening up more of the stack behind it — high-performance attention kernels, MoE communication library, and infrastructure for running agent environments at scale.
Model weights: huggingface.co/moonshotai/Ki…
Tech report: github.com/MoonshotAI/Kimi-K…
Tech blog: kimi.com/blog/kimi-k3
Some goofy rumors that Karpathy just left Anthropic today.
Maybe he joined to get the latest model recipe and we'll finally get this video.
Did Karpathy remove Anthropic from bio? Or was it not there.
Readers added context they thought people might want to know
Karpathy denied rumors of leaving Anthropic, calling them "weird misinformation." x.com/karpathy/statu…
People love to meme that the RL env space is oversaturated, but it really does makes sense why there’s a zillion data vendor startups.
For labs, only a very small team controls experiments on model arch, infra, and learning algos. And even then it feels somewhat limited: there’s only so much tinkering you can do with IS clipping, length penalty, async staleness, etc. Most tricks are known and is just a matter of implementing rather than experimenting.
Everyone else, which pertains a much higher headcount, is essentially data + evals. Models are always hungry for harder tasks and new verticals. Many of the vendors I’ve talked with in my job search mentioned they don’t have the capacity for all the requests they want to take. That’s why new data startups keep blooming.
TPOT is the most supportive environment I’ve ever been in. The AI/research/startup community is incredible, and everyone is so quick and gracious to offer a helping hand.
Within just a day of our entire org being cancelled, I’ve already been outreached by many inspirational researchers and engineers I would’ve never dreamed of meeting.
It’s been sad to talk about recent news, but also fun to discuss insights about CUA, infra, data, and evals with fresh perspectives. Onto the next.
Yesterday was my 23rd birthday.
As a present, I got laid off from Amazon AGI!
I’m actively looking for new opportunities, particularly related to computer use agents and post-training.
Since October 2024, I've been working on GUI agents non-stop: computer use at @SimularAI, mobile use @Apple, and browser use @AmazonScience.
With recent benchmarks like OSWorld 2.0, ALE, etc., it's clear CUA is far from solved, and would love to continue pushing.
Please reach out if you have any openings!
Yesterday was my 23rd birthday.
As a present, I got laid off from Amazon AGI!
I’m actively looking for new opportunities, particularly related to computer use agents and post-training.
Since October 2024, I've been working on GUI agents non-stop: computer use at @SimularAI, mobile use @Apple, and browser use @AmazonScience.
With recent benchmarks like OSWorld 2.0, ALE, etc., it's clear CUA is far from solved, and would love to continue pushing.
Please reach out if you have any openings!
Now it feels distillation is not about “uplifting your model higher”, but rather “uplifting your model faster”.
The advantage of distilling is you don’t need to spend a lot of time iterating on data synthesis for new verticals. Collecting trajectories on a new vertical which no model is good at is hard.
For example, there was some point in time 2 years ago when LLMs were bad at excel spreadsheet generation. Being the first lab to create 100k excel trajectories using the then-niche openpyxl library requires lots of thought and iteration on data generation. How can one generate high yield and high quality trajectories when the teacher model itself is not well-verse in this new vertical?
But, once another lab does all the hard work for you, you can just distill all that knowledge. Easy peasy. Especially if resource constrained, one can just wait for another lab to do the dirty work of bootstrapping knowledge.
Where I currently stand on how distillation is used and what performance uplift it uses.
For context, Anthropic HAS said that DeepSeek and others are using their models in an RL shaped data pipeline, but that does not mean it has substantial effect. These were very small numbers of samples, and likely to initialize an internal model or do a small training experiment.
The place where distillation is used as a key step is in SFT and/or midtraining (for seeding reasoning behaviors, thinking of sft and midtraining as separate is not helpful). This is why the Chinese models say they're claude -- when training with next-token prediction, the models learn "features" that are clusters of tokens, which will regularly appear in model outputs.
This SFT data is generated by skirting the intended behavior of the API and getting reasoning tokens out. Without getting the reasoning tokens directly from the model, the data would be very hard to train on (and likely not include the "I'm claude" behaviors).
SFT datasets have been O(1 million prompts) with high quality completions. Today, getting the prompts is often the hardest part -- especially prompts with carefully designed environments for proportional RL later, but this is mostly about SFT.
Getting to the end of it. The part where this SFT distillation is likely very useful is when a model moves into a new domain, a great example being the CripPT physics benchmark, where the US labs are far ahead. The idea would be to get some prompts, and get completions from a frontier model that you can use for an initial SFT.
With an initial SFT set, there is A TON of work to still do to get a model in the performance ballpark of GLM 5.2 and Kimi K3. The SFT stage is the start of a long, strenuous process for building a post-training recipe. It involves generating more SFT data and filtering it, something like rejection sampling, and polishing the behavior with extensive RL.
For models like Kimi K3 and GLM-5.2, the timeline is such that Fable 5 likely had no impact as a distillation teacher. That could help future models, though.
And as post-training becomes more dependent on techniques like multi-teacher on policy distillation, rather than just RL, there is even more complexity in how SFT distillation helps the final model. There are a lot of steps, and RL has been meaningfully scaled up in it's proportion of the final performance.
The real kicker for this is that in the above SFT stages, the best teacher models aren't normally the easies to integrate into the post-training recipe! The leading fully open, reasoning SFT works (open thoughts and olmo) have had a very hard time updating their recipes to use the strongest models as teachers. Often it is a smaller, surprising model which is the best teacher. This means that it may not even be the cutting edge models that are enabling distillation! Messy.
All together, the evidence in post-training is that distillation is becoming less impactful. Previously, before scaling RL, distillation was more impactful because the relative amount of performance gained from SFT on top of the base model was far higher.
I expect this trend to competitive, with how strong the best open weight models are -- and the teacher ambiguity clamps down on arguments that open-weight models are just "distillation washing" by removing the need to use Claude/GPT APIs. It could be that open models are just genuinely easier to distill from, by being easier to modify and tinker with.
The most impressive part of this announcement is NOT "second only to Fable".
Rather, it's the small detail that the model is only titled "Qwen 3.8".
One would think scaling params by 6x (from 397B to 2.4T), as well as beating GPT 5.6 and friends would warrant a bump up to "Qwen 4", but clearly not.
Hopefully, this means even stronger Qwens that will be an even bigger step forward are already cooking.
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
international：qwencloud.com/pricing/token-…
China：platform.qianwenai.com/prici…
Uhhh… 99% on HLE?
What’s going on here? MFW model reward hacks and search tool finds the golden answers online?
At first i thought this was a prank but apparently they have a whole tech report. Pretty committed for a prank.
Introducing Monolith-1.0: Frontier Reasoning, Open Weights
- 1.57T-param Mixture-of-Experts, 49.5B active per token
- Native 1M-token context (2²⁰), extended via a two-stage YaRN curriculum
- Trained on 60T tokens across 12,288 Ascend 910C NPUs
- Frontier SOTA open-source model as of July 2026: 95.9% GPQA Diamond, 96.2% MMLU-Pro, 90%+ AIME 2025
- MIT-licensed weights, tokenizer, and eval harness
No waitlist - open the chat and give it something that should be hard.
Chat: basaltlabs.org/chat
Readers added context they thought people might want to know
Independent analysis of the model weights indicates it is fabricated from Qwen2.5-7B-Instruct with cyclically reused and random-padded weights, not a trained 1.57T model. x.com/sshthedev/stat… huggingface.co/basaltlabsai/m…
The "distillation attack" narrative is FAKE.
Kimi K3 is so good because this distillation narrative has always been silly. RL is a bigger moat for getting this impressive level of hillclimbing; Supervised learning alone (i.e., distillation) does not take you this far.
Distilling from Teacher X + doing RL on higher quality tasks on a particular domain can outperform Teacher X on that domain. RL lets you reach the edge or even surpass your supervised barrier; This is not a new finding.
Currently, those "distillation attack" believers are thinking: "grrr... this Chinese lab is distilling and copying agent traces from this really good American model"
Instead, they should be realizing: "grrr... this Chinese lab is buying data and copying data synthesis methods from this really good American RL vendor"
Of course, both are chud mentalities to have. Incredible to see open weights continue to achieve the frontier.
Excited to see a solid open weights model from Thinky!
Also interesting to see more labs being very open about length penalty analyses, joining the likes of Cognition, MAI, Muse Spark, and Chinese labs.
Feels like RL leading to bimodal CoT lengths via hyper-terseness (lacking complete sentences, lacking grammatical glue), or hyper-yapping (spamming fork tokens like 'wait', 'actually', 'however', etc.) is under-discussed, even though it seems many labs runs into this.
Today, we are introducing Inkling.
Inkling reasons efficiently across text, image, and audio modalities. We are making the full weights available.
thinkingmachines.ai/news/int…
Available today for fine-tuning on Tinker. Play with it in the Inkling Playground. 🧵
Recently, I've been revisiting OSWorld and looking through the trajectories of frontier models. Last time I did such analyses was over a year ago when SOTA was low 40s.
Back then, I expected most of the remaining head room would come from better exploration: Through RL, your model becomes good at meta-behaviors like reflection + backtracking, which helps the model explore complex apps and find the right widget/interface to use.
However, it couldn't be further from that case. Frontier trajectories exhibit much less exploration than I expected. For example, models will somehow magically beeline to the correct nested submenu, or perfectly interact with complex chart configurations.
The biggest improvement don't seem like improved reasoning/meta-behaviors. It just seems like these models have much more knowledge about every single menu and widget than before.
