Hen Sapir

@amanrsanger wait, CPT on an instruct model?

We've evaluated a lot of base models on perplexity-based evals and Kimi k2.5 proved to be the strongest! After that, we do continued pre-training and high-compute RL (a 4x scale-up). The combination of the strong base, CPT and RL, and Fireworks' inference and RL samplers make Composer-2 frontier level. It was a miss to not mention the Kimi base in our blog from the start. We'll fix that for the next model.

@SeanZCai @PrimeIntellect if you have ground truth samples, use gepa to optimize your judge (rubric, criteria weights, prompt). cc @JoshPurtell who built docs.usesynth.ai/cookbooks/veri…

Running @PrimeIntellect Lab GRPO on a hard-to-verify action-matching task. Judge inconsistency was generating phantom reward variance like same model output, different scores across rollouts and step 0 kept winning. Fixed it with a stronger judge (better SOTA OAI model, using all my thousands of old OAI hackathon credits) + response caching and got zero phantom variance, clean gradient signal. Anybody know what's the best off-the-shelf judge for semantic action matching in RL training without post-training a purpose-built one? What are people actually shipping with? Is there anybody working on purpose-built judge models for GRPO?

@jeff_weinstein @sumeetvtweets

who is building the next great ci startup optimized for agent generate code? (there is about to be _extremely_ ci needs)

@marcklingen @langfuse request to add the ability to set negative tag filters in the trace UI (e.g., tag != X). i can hack it via query params but would be nice to have it in the UI 🙏

@jaltma any company whose core value prop is accountability

The current consensus view is saas is dead...presuming that's right, the next interesting next question is What companies are "safe from ai"? - handling money, regulation - agents on top of company data - most hardware? - maybe systems of record? - security? - marketplaces?

@jarredsumner how would this would work with encodings? you can't decode partially-encoded characters that have been cutoff by maxLength/offset. and, afaict, all workarounds to that are bad

node:fs readFile needs a maxLength and offset option

@irl_danB interested!

I have a working version of the call stack context manager as an opencode plugin. It exposes tool calls to the agent to manage the call stack as plugin state. Frames in the call stack are opencode sessions. Compacted ancestor, uncle, and sibling frames are injected into the context. I’ve not run benchmarks and it needs plenty of tire kicking, but opus 4.5 uses it pretty well now. It’s not really built for interactive mode at the moment (it can be), I haven’t optimized it for cache utilization (versions of it can be), and for now it is most useful for certain classes of tasks: primarily long running, like building out a large new project or doing a large refactor. In fact adding this plugin kind of borks your opencode if you’re trying to run short single threaded interactive tasks, because it aggressively uses the call stack model to break apart tasks. It ships without expectation of maintenance or further attention, so don’t build anything on top of it. Normally I wouldn’t release this at all, except several people have indicated interest and willingness to play with a raw version of it. If you want to play with a more polished, actually useful, hopefully benchmarked version, please wait a few more days. Please forgive the slow progress, I’ve got my hands full at the moment with two newborns and two toddlers. I’ll run it on terminal-use and harbour benchmarks once I shore up confidence that it’s worthwhile to spend that money. Buried lede: building this out has made me realize that the opencode session itself is a unique primative that can probably support more interesting composition. Pair this with the opencode client-server model and I’m thinking of pivoting my currently in-progress inversion of control framework from orchestrating Claude Agents SDK to orchestrating opencode sessions. Exciting stuff ahead, lots you can do with this. Please reply or dm if you are interested in trying the proof-of-concept version, especially if you’re interested in sending feedback. Wait a bit if you want something more polished and proven.

what a gift 🇺🇸

@corbtt @willccbb do you think that will still be true in 1yr? presumably not

@willccbb 10B+ is a pretty low bar. Most actual usage of open models is in the 100B+ range.

registering a prediction that by this time next year, there will be at least 5 serious players in the west releasing great open models chinese models will still be great, but the gap will be small if not non-existent, and people will mostly use the western ones

@nikitabier @dinkin_flickaa @misha_mityushk @nicoduc ah okay i see. wasn’t immediately obvious but will get used to it 👌

@hensapir @dinkin_flickaa @misha_mityushk @nicoduc You can swipe right to get back to timeline

We're testing a new link experience, starting on iOS -- to make it easier for your followers to engage with your post while browsing links. For creators, a common complaint is that posts with links tend to get lower reach. This is because the web browser covers the post and people forget to Like or Reply. So X doesn't get a clear signal whether the content is any good. To help get better signal, posts will now collapse to the bottom of the page so people can react while you're reading. As always, remember: the post should stand alone as great content so write a solid caption.

@nikitabier @dinkin_flickaa @misha_mityushk @nicoduc one unexpected annoyance is that, when exiting the browser view, i get redirected to the post detail view (ie the /status/<post_id> page) even if i clicked the link from my home timeline. otherwise, nice work 👏

Credit to @dinkin_flickaa and @misha_mityushk for building it and @nicoduc for the designs. It's only version 1, so please share any bugs you find.

@simonw x.com/cognition/stat… related - presumably this is rl-finetuned on a small oss llm but the blogpost doesn't confirm/deny that hypothesis. cc @cognition

@simonw also lots of examples where the dev time costs are meaningfully lower, ie prompt iteration on frontier models taking longer than RL-fine tuning small oss models for the same task

Anyone got a success story they can share about fine-tuning an LLM? I'm looking for examples that produced commercial value beyond what could be achieved by prompting an existing hosted model - or waiting a month for the next generation of hosted models to solve the same problem

leggo

@growing_daniel @tigran_zzz it’ll probably be a net positive but america “lacks” israel’s mix of existential threat, national cohesion, tiny scale, and integration of army experience into daily life. i think all of the above are required for the outcomes you’re thinking about.

@tigran_zzz Israelis famously underperforming Maybe your country just sucked

Two years of military service for everyone after high school or turning 18 would fix America

@sentdefender you already know this but… in the middle east, there’s what you say, what you do, and what you think—and none of them are ever the same.

Despite U.S. President Donald J. Trump stating yesterday that Saudi Arabia was now willing to “Normalize Ties” with Israel without the guarantee of Palestinian Statehood; the Kingdom of Saudi Arabia release a Statement earlier, stating that this was not true, and that there would be No Diplomatic Ties with Israel unless a Palestinian State is established with East Jerusalem as its Capital.

@andrewchen @KatiaAmeri i'm biased but @FrontHQ wins by a landslide front.com

dear lazyweb- what's everyone's new modern customer support tool? (no more zendesk!!!) cc @KatiaAmeri

@DD_Geopolitics hamas-massacre.net

@soffes You can create a Twilio SMS team inbox with @FrontApp! See help.frontapp.com/t/q52442/how-t…

Looking for a service that gives you an SMS number and goes to a team inbox to use for support. Would love for users to be able to SMS us for help. Know of anything like that? (I know I could make something but would rather pay for a service that does this well.)