Chip Huyen

How Jev-style decision models can make agent memory smarter Jev has been everywhere lately. In short, it's a model that doesn't write text. It picks the best answer from a list you give it, and does it fast and cheaply. That got me thinking about agent memory, which is mostly small decisions like that. Right now we make them with rough rules or pay for a full LLM call each time. Some places I think a decision model could help: - Is this conversation worth remembering, or is it just chatter? - Is this new info an update to something we already know, or something new? - Are these two names the same thing? - What kind of search should this question use? - Is the agent about to do something risky, and is it worth stopping to check? In Memoose, my open-source memory tool for agents, I've done some early work with Laya, an open model in the same style as Jev. It got good results on a few actions: deciding what's worth remembering, spotting when the agent is about to do something risky, and picking the right kind of search. It runs locally, with no API key. It's still early, so I'd love for you to try it out and tell me what you think. https://lnkd.in/gNq455Zr

24

1 Comment

I’ve been following ICSE since I was a PhD student, and it’s wild to see how much the conference has grown. ICSE 2026 (Research Track) reported 1,469 submissions with 321 acceptances (~19.8%). Back in the mid-2010s, the numbers were a lot smaller—e.g., ICSE 2016 had ~530 submissions and 101 acceptances (~19%). Same elite selectivity, but on a much bigger scale. Beyond the raw volume, the topic mix has expanded too: AI for SE / SE for AI, large-scale testing & analysis, security and trustworthiness, DevOps/release engineering, and empirical studies at massive scale. It’s a good signal of how central software engineering research has become—and how much the field is being reshaped by real-world complexity (and now LLMs). On a side note: the most recent completed CVPR (CVPR 2025) had - 13,008 submissions - 2,872 accepted (≈ 22% acceptance rate) #ICSE #SoftwareEngineering #SEResearch #AcademicTwitter #PhDLife #ResearchCommunity #EmpiricalSE #DevTools

13

1 Comment

I wrote a summary of 𝗠𝗮𝗴𝗶𝘀𝘁𝗿𝗮𝗹 ✨, Mistral AI's first reasoning model The standard for giving a model "reasoning power" is: • Supervised Fine-Tuning (SFT) -- a model learns the reasoning "language" • Then, Reinforcement Learning (RL) -- improve such reasoning However, what's cool about Magistral is that the model learns to reason *only* from RL Why this is important is RL allows the model to *actually learn* reasoning, whereby SFT trains the mode to *copy* a reasoning style, often from a stronger model Notably, DeepSeek already tested this "RL-only" hypothesis: • Their model did learn to reason and do math very well • BUT the model also had issues with multilingual (mixed languages randomly) & poor formatting • After all, the model didn't have SFT to imitate a reasoning "language" and instead had to figure out reasoning by itself Mistral fixed these issues in their 𝗥𝗟 𝗿𝗲𝘄𝗮𝗿𝗱 𝗱𝗲𝘀𝗶𝗴𝗻, steering a model's behaviour by giving rewards for specified outcomes If you're interested in how they designed the reward, read more to find out :) It was an enjoyable read! Summary link in comment 👇 P.s. I thought "Magistral" comes from the word "Magic" because of their similar pronunciations, hence the magic-vibe background -- Turns out, it doesn't lol.

16

4 Comments

What actually happens inside an AI agent between receiving a request and producing a response? Not the topology. Not the arrows between boxes. The engine inside each box. I mapped the 5 internal subsystems every agent needs — and the design decisions that determine whether it works in production or just in demos. ──────── Spoiler: context assembly (how you build the prompt at each step) matters more than which model you use. ──────── 👉 Read on Medium: https://lnkd.in/gTKp-2Sk 👉 Read on Substack (free but the same content on Medium): https://lnkd.in/gGHK2KpA  #AIAgents #SoftwareEngineering #AI #LLM #AgentArchitecture

37

It's a humbling moment to share that our paper has been accepted at NeurIPS 2026, one of the world's top AI conferences (CORE A*). This year, 7,900 of 30,709 valid Main Track submissions made it (25.7%), ours among them as a poster. 🎉 🔍 Time series foundation models (TSFMs) are often adapted with a mixture-of-experts head: a router plus specialist experts. In finance, a calm week and a volatile spike call for different treatment. Yet on many of these models, the router collapses once the backbone is fine-tuned too, and one expert takes every input. 🎯 Our paper is about knowing where to strike. Some TSFMs normalize each input window, erasing the level and scale a router needs to tell a calm week from a volatile spike. Our fix, Raw-Routed Mixture of Adapters (RR-MoA), changes one thing: the router reads the raw input, and routing recovers. 📈 The payoff: RR-MoA wins 54 of 54 comparisons against the strongest fixed adapter, and with a frozen backbone it even beats full fine-tuning. The paper has the full story, including the eight controls behind the causal claim. 🙏 Huge thanks to my trusted co-authors Dr Thuy T. Nguyen, Dr Minh Ngoc Dinh and Dr Nhat-Quang Tran. See you in December at NeurIPS! #NeurIPS2026 #TimeSeries #FoundationModels #MachineLearning #Finance

54

4 Comments

Bonsai 1.7B ternary model at 442 tokens/s on M4 Max. Last week the same engineer built an end-to-end 3D reconstruction pipeline. Today, model deployment. PrismML recently released a family of efficient 1-bit and ternary LLMs called Bonsai, which we believe is a big step forward for deploying AI everywhere. We took the Bonsai 1.7B ternary model and ran our agentic evolution search on it for 6 hours to optimize Metal kernels. Measured against unmodified upstream llama.cpp at the same Bonsai/Q2_0 commit, same M4 Max: - tg128: 309.82 → 442.42 t/s (+42.0%) - pp512: 4250.32 → 4622.63 t/s (+8.8%) We saw similar speed-ups across Apple A chips, Qualcomm Snapdragon, NVIDIA GPUs, and x86. Try it here: https://lnkd.in/gNkJPc9V #Apple #NVIDIA #Qualcomm #ARM #Intel

8

gave a short talk on my thoughts on multi-agent systems. talked about ways to run more agents in parallel, how to run agents for longer, and some forward looking thoughts on how this space evolves. https://lnkd.in/gVfGJbcE

11

1 Comment

A New Kind of AI Reasoning — And Yann LeCun Is On Board A 6-month-old Silicon Valley startup called Logical Intelligence just unveiled something different: an energy-based reasoning AI named Kona. Instead of predicting the next word like LLMs, Kona scores entire solutions against constraints and searches for the lowest-energy (most consistent) answer. Less guessing. Fewer hallucinations. Big signal: Yann LeCun has joined as founding chair of its Technical Research Board — a strong vote of confidence in this direction. Kona is non-autoregressive. It generates full reasoning traces at once, then iteratively fixes weak points by minimizing an energy function. Think optimization, not storytelling. The model operates in a continuous space, allowing precise edits without regenerating everything — making it more efficient and potentially far more reliable for math, logic, and proofs. Logical Intelligence is now raising at a $1B–$2B valuation, betting that energy-based models are the path beyond LLM limits. This is very much LeCun’s long-standing thesis: AI shouldn’t guess the next token — it should optimize toward truth. https://lnkd.in/gbSD4d89 : The probability of finding the particle at a given point x is spread out like a waveform, there is no definite position of the particle. Everything are probabilities.

17

In the past week, two serious open-weights contenders dropped: 𝐆𝐋𝐌 𝟒.𝟕 𝐌𝐢𝐧𝐢𝐌𝐚𝐱 𝐌𝟐.𝟏 What’s striking isn’t just that they’re “good for open models” — it’s that across coding (SWE-bench), math (AIME), and general reasoning (MMLU, GPQA), the gap with top closed labs is now measurable, not structural. In some slices, it’s already razor-thin. This signals a clear strategy from China: release a strong base model, then let startups, researchers, and enterprises fine-tune, self-host, and compound value on top. Closed labs still lead at the frontier — but open weights are no longer chasing from behind. They’re compressing the distance, FAST!

8