Ask stupid question-Think impossible-Do step by step.
Joined January 2023
- Tweets1.1K
- Following906
- Followers17
- Likes3K
asfaan murthyulas retweeted
Very important work. One of the things missing from local learning (PC etc) is fast weights. It is missing no more!
The connection to deltanet follows so naturally here.
asfaan murthyulas retweeted
科研 Agent 会提 idea 已经不稀奇了，难的是知道哪条真的值得继续做。
这篇 DiscoveryRSI 想让 Agent 学会筛选和改进研究假设。
它会围绕一个假设不断生成、修改和组合方案，同时要求每个假设都有明确证据、实验设计、判断标准和替代解释。
更关键的是，研究过程会反过来训练下一轮 Agent。
好的假设会保留下来，失败轨迹则用来定位研究流程哪里出了问题，再继续更新模型和 Harness。
两轮自我改进后，假设质量从 4.86 提升到 5.39，提升主要来自「可证伪性」和「数据是否真的能回答问题」。
还有一个挺有意思的结果：
DiscoveryRSI-27B 生成的假设中，有 6.3% 和模型知识截止后发表的研究方向相似；GPT-6 是 2.5%，Claude Opus 5 是 2.3%。
这个指标当然不能直接等同于真正发现，但方向很值得关注。
科研 Agent 的自进化，不该只学会写得更像论文。
它还要逐渐学会判断，哪些问题值得追，哪些证据真的够，以及下一轮研究该怎么做得更好。
📎 arxiv: arxiv.org/abs/2610.23839
asfaan murthyulas retweeted
1/ The tokenizer is a new scaling axis!
We introduce an approach for diffusable representation learning that's principled, scalable, and works when DINO/LeJEPA collapse.
We get joint tokenizer+diffusion scaling laws and it's compute-optimal vs raw diffusion
w/ @drfeifei @GordonWetzstein @jcjohnss @ericryanchan @vincesitzmann @KyleSargentAI
Thread, Paper, and Blogpost below
asfaan murthyulas retweeted
A looped transformer has to know when to stop, and its state has to stay bounded while it loops.
We introduce Spherical Fixed-Point Reasoners (SFPRM) 🌐: loop on the unit sphere, where each step is an angle, and stop once the angle is small enough.
🧵 (1/9)
asfaan murthyulas retweeted
“We found DeepSeek V4.1 Flash 1017 beats GPT 6 Sol on our multi-step offensive cybersecurity task.”
READ THAT AGAIN.
the same cyber capabilities that OpenAI kept behind a gate got beaten by a Chinese open model you can run locally.
buckle up guys.
We found DeepSeek V4.1 Flash 1017 beats GPT 6 Sol on our multi-step offensive cybersecurity task. Sol was tested via the Daybreak trusted access program.
We have been building a cybersecurity benchmark based on simulated corporate networks. This is a valuable indicator for the cyber-risk a model poses if in the hands of a bad actor. We plan to make the full evaluation public in the coming weeks but wanted to share our first immediate finding here.
asfaan murthyulas retweeted
🚀 Introducing DeepInfer, our in-house inference framework, now fully open source!
Powering DeepSeek-V4.1-Flash at scale with high throughput and low latency. We’re sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable.
🔗 github.com/deepseek-ai/DeepI…
1/4
asfaan murthyulas retweeted
多様な提案を進化させ、意思決定を支える推薦手法「QD-Rec」を開発
ブログ：sakana.ai/qd-rec/
論文：arxiv.org/abs/2610.23470
GitHub：github.com/SakanaAI/QD-Rec
どの企業を買収候補として検討するか。どの顧客に新しい製品を提案するか。膨大な論文の中から、次に何を読むか。こうした意思 決定では、候補をスコア順に並べるだけでなく、「なぜ選ぶのか」「選んだ先にどのような可能性があるのか」を、異なる観点から比較できることが重要です。
一つの候補にも複数の有望な理由が考えられます。数千件、ときには1万件を超える候補それぞれについて、さまざまな仮説を生成・評価し尽くすには大きなコストがかかります。私たちは、限られた予算の中で「候補」と「仮説」の組を探索し、質が高く多様な提案を見つける問題を「オープンエンドな推薦」として捉えました。
QD-Recは、オープンエンドな探索の研究で発展してきた「Quality-Diversity（QD）」と呼ばれる進化的探索技術を、推薦に応用します。過去の評価結果をもとに次に調べる候補を選び、蓄積された仮説の着眼点をLLMが別の候補に応用したり、複数の仮説を組み合わせたりして、新しい提案を生成します。新しい切り口の提案は残し、似た提案同士ではより優れたものに更新することで、多様な仮説を次の発見の「踏み石」として活かし、推薦の集合を進化させていきます。
実験では、 企業の買収候補探索、架空の顧客プロファイルを使ったターゲティング、研究者向けの論文推薦などで検証しました。同じ評価回数のもとで比較したところ、多くの設定で、比較手法よりもQD-Recの方が評価が高く、より多様で多くの推薦を蓄積できました。
私たちが目指すのは、AIが人間の意思決定を代行することではなく、多様で有望な選択肢と、それらを比較・検討するための判断材料を提供することです。今回の成果を出発点に、QD-Recを次世代の意思決定を支える基盤技術へと発展させていきます。
asfaan murthyulas retweeted
同じ1つのニューロンで、ニューラルネットワークの計算を再現できる。
そんな衝撃的な研究がarXivに公開されてたんだよね。
ハーバード大などの研究者による論文
「The Surprising Expressivity of a Single Neuron with Autoregressive Chain-of-Thought」
一般的なAIはたくさんの部品を組み合わせて複雑な計算をしています。
でもこの研究では、たった1つのニューロンでも、何度も計算を繰り返すことで、ニューラルネットワークの計算を再現できると示しています。
もちろん、これでLLMなどのAIモデルを劇的に小さくできる！っていうことじゃありません。1つで計算する分、何度も繰り返す必要があります。
それでも、
自分が面白いと思ったのは、AIに複雑な能力を持たせる方法って、思っていたよりいろいろあるのかもしれない、ということ。
今まで当たり前だと思っていたAIの作り方を、考え直すきっかけになる研究。
論文：arxiv.org/abs/2610.17283
asfaan murthyulas retweeted
Two years of work, and PySR v2 is finally out!
If you haven't used it before: PySR is a machine learning library that finds short equations that fit your data.
I've always wanted a PyTorch for symbolic learning. PySR v2 is a big step towards that:
asfaan murthyulas retweeted
Chinese researchers just published a paper with a brutal title: "The End of Software Engineering."
their argument: writing code is finished as a skill
a human brain can juggle only so much state before it locks up
an agent hits no such wall, and more compute keeps raising its limit
typing stops being your part of the job
you become the intent architect: you decide what gets built and sign off on what comes back
the next raise goes to whoever can tell a solid agent run from a confident wrong one
so i wrote the guide below on building an agent system of your own
the paper behind it is in there too
asfaan murthyulas retweeted
An Anthropic researcher just told a room full of PhD students that their career plan no longer exists
His 3 pieces of advice for students:
1. Use AI every day. If you're on the exponential, you'll never fall behind it.
2. Choose projects that get better when compute scales. The rest will be solved before you finish them.
3. Stop planning your career as if it were stable. The safe option no longer buys safety; it no longer exists.
And a phrase that sums it all up:
"Writing is thinking. If you outsource your writing, you outsource your thinking."
The talk is titled "The End of the Anthropocene"
asfaan murthyulas retweeted
Can research agents get better at science by learning from the research they do?
Introducing DiscoveryRSI 🧬: a framework that turns hypothesis-driven research into a recursive self-improvement loop.
📄 Paper: arxiv.org/abs/2610.23839
💻 Code: github.com/Gen-Verse/Discove…
🔴 Live Arena: foundry-arena.rsi-arena.work…
🔹 What is a valuable hypothesis?
We define it as a joint object H = (claim, investigation plan, decision criteria, alternative explanations). A hypothesis is valuable if resolving it would change a research decision. It is testable if the plan can produce evidence that supports, weakens, or bounds the claim.
🔹 How it works
- Lead, Generation, Evolution, Critique and Progress agents grow hypothesis trees through evolve / repair / crossover / restart
- Every version stays linked to its evidence, data, assessments and contribution records
- Selected plans compile into executable Harbor tasks
🔹 Three self-improvement mechanisms
① Hypothesis credit optimization: blinded TrueSkill assessments assign bounded credit to the agent steps that actually contributed, and DiscoveryRSI-27B is trained only on those steps (Qwen3.8-27B init, 256K context; 2,397 of 15,792 trajectories kept; 2.12% of tokens supervised)
② Harness replay: diagnose recurring failures such as repeated exploration, failed tool use, lost artifacts and premature closure, then refine the workflow
③ Rediscovery: run the accepted model and harness updates on fresh questions
🔹 Results (DFM-Bench, 4 life-science domains)
- Hypothesis quality 4.86 → 5.18 → 5.39 over two rounds
- Falsifiability 4.00 → 6.20, alternative explanations 4.44 → 5.98, data resolvability 3.73 → 4.99
- Evolve/repair share 77.5% → 91.7%, restarts 21.7% → 5.0%: the agent learns to keep developing an idea instead of starting over
- Caveat: importance and novelty dip slightly. The model learns to design better tests, not yet to choose bolder questions.
🔹 Expert review
27 physicians and bioscience researchers scored 292 anonymized proposals:
- 41% of DiscoveryRSI-27B revisions improve quality vs 31% (GPT-5.6 Sol) and 25% (GPT-6 Luna)
- Only 12% make things worse vs 32% and 31%
🔹 Does it anticipate real science?
Against 16,440 publications filtered by each model's cutoff, 6.3% of DiscoveryRSI-27B's hypotheses connect to later published research (16.9% in therapeutic targets). Comparison: DeepSeek 5.4%, GLM 2.8%, GPT 2.5%, Claude 2.3%.
Watch six models do research in real time in DiscoveryRSI Arena, and follow every proposal, critique and compile step 👇
foundry-arena.rsi-arena.work…
Co-led by Jianyuan Zhong, Chenwei Wu & Zhaochen Yu, with wonderful collaborators across PhAI Labs, CUHK, Stanford, MIT, Harvard Medical School, Oxford, Princeton, PKU and more.
Grateful to @james_y_zou @Jiaru_Zou @Charles_Y_Wu @zhenfei_yin_yzf 🙏
asfaan murthyulas retweeted
This is f*cking gold..
Chinese researchers just dropped a paper called "The End of Harness Engineering" and it reads like an obituary for LangChain.
The claim: the file that organizes the model should be written by the loop that runs it, not by the person who built it.
A hand-built harness is a guess with a commit message. A self-revising harness is a file that grades itself every round.
The mechanics are almost boring. Execution report in, 8 edits out, score them on held-out questions, keep the winner, repeat. Frozen solver the whole time.
Across 41 rounds the held-out score went from 0.31 to 0.64 and the harness never saw an editor.
The human writes exactly one thing: the scoring set. Never shown to the proposer, never optimized against, used to grade every candidate.
So you stop arranging calls and tools. You start deciding what counts as correct and reading the score.
Harness engineering turns disposable. The scoring set turns into the job.
That is exactly the loop I run and documented below: one harness.py, 300 Kimi K3 workers, 37 rewrites in 41 rounds, zero edits by hand.
The whole paper in one line: grade the file, don't write it ↓
asfaan murthyulas retweeted
Introducing Optimus-1, the first swarm harness for autonomous research. Generally available in ArkenOS, starting now.
Optimus-1 lets ArkenOS spawn a swarm that runs the entire discovery loop on hard research problems autonomously. The harness samples the hypothesis space in parallel and works the tails of the distribution, chasing every anomaly until it resolves into an artifact or a discovery. We want every researcher to have a lab of their own, and to see how far their curiosity can go.
Unlike coding agents, Optimus-1 is designed for long-horizon discovery, where most branches fail, and the hard part is knowing which result to believe. It keeps what replicates and keeps pushing past the edge of what's known.
We're still early, but ArkenOS is helping hundreds of teams, from RL environments and data labs to frontier research groups, scale the alpha in their core research.
If you're working on a hard research problem and want to try it, my DMs are open. I'd love to hear what you'd point it at.
asfaan murthyulas retweeted
Looped Models also overthink!
Just like reasoning LLMs burn tokens on problems they’ve already solved, looped Transformers keep spending loops after the answer has settled. Much of the computation across loops is redundant, and sometimes even harmful.
The takeaway isn’t “loop less.” It’s “loop smart”: residual connection is the key for looped models. Check out our paper: arxiv.org/pdf/2610.11570.
Looped Transformers should get better the more they loop. In practice, more loops often make them worse. We call this latent overthinking.
We found why, and how to fix it. Introducing InfiLoop 🧵
📄 arxiv.org/abs/2610.11570
asfaan murthyulas retweeted
Google just open-sourced Harness².
(turn any harness into a self-improving system)
a harness is everything wrapped around the model, like the system prompt, tools, memory, skills, and subagents. today a person designs it once, and the agent runs every new task with that same fixed setup.
Harness² lets the agent refine that setup for each new task on its own, with the model weights left untouched.
on every task, the agent proposes edits to its harness, runs a few candidate versions, and composes a better harness by comparing how those runs behaved.
because it learns from the differences between runs, it needs no correct answer or grader for the task.
that covers a single task. Harness² also carries what it learns from one task into the next.
after each task, it saves the harness it settled on along with the lessons that held up across the runs. the next time it refines a harness, it starts from those notes instead of from scratch.
this means every finished task makes the next round of edits more informed.
Google tested this with Gemini 3.5 Flash and GLM-5 on legal, office, and professional work benchmarks. one round of edits was enough to beat their original harnesses, and giving the agent more rounds pushed scores higher.
those extra rounds can go wide or deep. the agent either tries several candidate harnesses at once or keeps refining a single one step by step, using OpenCode or Codex to carry out the work.
you can find the Harness² repo here: github.com/google-research/h…
I wrote about a similar idea for agents already running in production, where the harness should repair itself after a failure instead of waiting for someone to debug the trace by hand.
that piece walks through how @Cometml's Opik (open-source) closes this loop, going from a failing trace to a proposed fix to a regression test that stops the same failure from coming back. it's quoted below.
asfaan murthyulas retweeted
We plugged Google Research’s Harness² into JEV 1.14.
Zero weights touched, yet driving massive gains on vision benchmarks round after round.
Under the hood, the system:
⚡️ Rewrites its own prompts
🛠️ Builds its own tools
🧠 Distills its own experience
JEV-Harness² repo drops soon. Stay tuned.
Google just open-sourced Harness².
(turn any harness into a self-improving system)
a harness is everything wrapped around the model, like the system prompt, tools, memory, skills, and subagents. today a person designs it once, and the agent runs every new task with that same fixed setup.
Harness² lets the agent refine that setup for each new task on its own, with the model weights left untouched.
on every task, the agent proposes edits to its harness, runs a few candidate versions, and composes a better harness by comparing how those runs behaved.
because it learns from the differences between runs, it needs no correct answer or grader for the task.
that covers a single task. Harness² also carries what it learns from one task into the next.
after each task, it saves the harness it settled on along with the lessons that held up across the runs. the next time it refines a harness, it starts from those notes instead of from scratch.
this means every finished task makes the next round of edits more informed.
Google tested this with Gemini 3.5 Flash and GLM-5 on legal, office, and professional work benchmarks. one round of edits was enough to beat their original harnesses, and giving the agent more rounds pushed scores higher.
those extra rounds can go wide or deep. the agent either tries several candidate harnesses at once or keeps refining a single one step by step, using OpenCode or Codex to carry out the work.
you can find the Harness² repo here: github.com/google-research/h…
I wrote about a similar idea for agents already running in production, where the harness should repair itself after a failure instead of waiting for someone to debug the trace by hand.
that piece walks through how @Cometml's Opik (open-source) closes this loop, going from a failing trace to a proposed fix to a regression test that stops the same failure from coming back. it's quoted below.