AI is now improving its own model architectures and chips. The power infrastructure it runs on deserves the same attention. We used RL to teach a Kimi base model to design medium-power transformers. It met 93% of unseen specs, helping to compress multi-week engineering efforts into a few minutes of inference. 1/5

Using Tinker with an autoresearch loop is a really effective way to reproduce post-training papers at predictable costs Today there are dozens of self-distillation methods all claiming improvements over each other, and it’s hard to establish which claims hold up We gave agents a Tinker budget to reproduce self-distillation results across models and training setups. With just a few user prompts, they reproduced SDFT’s continual learning benefits across Qwen3-8B and Qwen3-30B-A3B over multiple seeds, and investigated SFT’s failure modes. Read more below:

How to post-train a model to personalize it on your code repo? In our latest research in Bespoke Labs, we post-trained a model to improve its performance on a given Github repository. Starting from Inkling base, we use supervised fine-tuning (SFT) with trajectories coming from a strong teacher model, and reinforcement learning (GRPO) on repository-specialized environments that we curated. SFT gave a 52pp improvement in performance on the held-out fontTools evaluation set. Further RL training lifts the total improvement to 57pp compared to the base Inkling model. In addition to the in-distribution evaluation our post-trained Inkling shows good performance on Terminal-Bench 2.1 and SWE-Bench Lite while becoming 40% more token efficient due to post-training. Read our full research blog post here: Many thanks to Thinking Machines Lab for their credit contribution that helped support this research.

New preprint from @lightningrodai, this time with Philip Tetlock and Ville Satopää! We post-train 5 versions of the same LLM, changing only the scoring rule used as the RL reward. Similar aggregate scores, but very different BIN profiles. A good Brier score alone doesn't tell you if a forecaster can distinguish likely from unlikely events. One might discern well but lose if its probabilities run systematically too high. A less discerning one might score better by hugging the base rate. BIN splits forecast performance into bias, information, and noise. Bias is a systematic shift in the probabilities. Noise is random scatter. Information is real signal about which outcomes are more likely — the part you need a powerful LLM for. Different uses call for different profiles. Reward choice is one lever shaping which forecaster you get. Congrats to co-authors @indiequant @KSkotheim64001 @VSatopaa @PTetlock 🙌 Full paper:

We published how @thinkymachines 's Inkling did on Voice Code Bench few weeks ago, and then we asked, what if we fine-tune Inkling? Inkling's main value prop is its small size and robustness, making it a good base model for domain specific fine-tuning. So we that's what we did. We fine-tuned the same Inkling speech model on 1, 25, and 100 hours of our proprietary data containing alphanumeric entities. And the curve kept moving. 🍊 Not surprising, on the standard 300-item VoiceCodeBench evaluation, the 100-hour checkpoint delivered the strongest result. Untrained Inkling → 100-hour post-trained Inkling: 📈 Task Success Rate: 56.33% → 79.00% (+22.67 pts) 📈 Entity recovery (CTEM): 86.84% → 94.80% (+7.96 pts) 📉 VoiceCodeBench WER: 2.3748% → 1.6107% (32.2% relative reduction) 🔧 139 misses fixed, 21 prior hits regressed: net +118 exact entities recovered And the gains scaled as data scaled: 1h: 88.06% CTEM / 59.33% TSR / 2.8188% VCB WER 25h: 92.85% / 72.67% / 1.8517% 100h: 94.80% / 79.00% / 1.6107% The 100-hour model recovered values the base model missed: •--revert-last, instead of splitting one flag into “--revert --last” • tests/auth/login.spec.ts, instead of test/auth/login.spec.ts • ALLOWLIST_CIDR, instead of inserting an extra underscore • SN-7KX-9042, instead of dropping the final digit The largest entity-type gains were email addresses (+30.8 pts), postal addresses (+30.0), file paths (+23.5), environment variables (+22.9), and IP addresses (+20.0). That’s the Besimple thesis: targeted human data can move the production metrics that matter for voice agents, even when the base model is already strong. DM me if you want to evaluate your model on this benchmark or build the data that moves it. 🍊 #SpeechRecognition# #VoiceAI# #ASR# #Transcription# #DataQuality# #PostTraining# #Benchmarks# Checkout the full blog at:

Introducing autoresearch for arXiv papers: replicate and experiment on any arXiv paper using an army of Claude/Codex agents For post-training, your agents can use @tinkerapi to launch concurrent RL runs while you watch your experiment tree grow Try it out now on OpenResearch:

Recently we are working on adding proactive support to the collaboration layer. While testing with different people, I found audio is a good way to capture moments when people need support. So we added a "Hi Coco" feature and support inkling model through @tinkerapi - it works pretty well! (video has sound!🔊) We open source this part first since several people told me they're exploring ideas around collaboration layers + talk-out-loud. Link in thread.

Introducing NVIDIA Nemotron 3.5 Lightning⚡ An open 30B MoE model with 3B active parameters, built for always-on agents to complete high-volume, specialized tasks faster. It delivers up to 4x the output speed of similar-sized models.

I RL-trained @thinkymachines Inkling model to tell dad jokes, all from my laptop. Inkling is a 975B open-weights MoE. I fine-tuned it with GRPO on @tinkerapi into PunTune-0.6, which answers any topic with a focused, original dad joke. The jokes are... fine. But the finding was insightful. The main discovery was that training didn't teach the model new wit. It taught it to pick one punchline... and stop. Something a lot of great models (and my mom) really struggle with. PunTune reaches strong punchlines on about a quarter of the tokens as the base Inkling model. My rewards bought reliability and economy, not necessarily creativity. This is the type of learning that helps product leaders start to connect how training, and understanding what is/isn't possible, can sharpen the value your AI models deliver. Especially when...

New research from Bridgewater AIA Labs, UIUC, and MIT: we prove what we believe to be the first non-vacuous generalization bounds for reasoning LLMs on real-world problems. RLVR powers frontier reasoning capabilities yet its generalization to unseen data has remained an open theoretical question and deployment blocker for practitioners. Our generalization bounds for RLVR deliver provable high-probability lower bounds of the accuracy for billion-parameter RLVR models on unseen data, which can provide guidance on safely deploying RLVR. 1/9