Juan Martín @IAconmate

Los datos que muestra @Hvygens 👇sobre evolución de la producción cientifica Argentina son devastadores. A fines de los 90, Argentina era líder científico indiscutido de América Latina: más papers y patentes per cápita que cualquier otro país de la región. Pero como en tantas áreas, el modelo estatista, proteccionista y burocrático destruyó la capacidad de investigar e innovar, y fuimos ampliamente superados por países vecinos. Impresionante: Chile hoy publica más del doble de artículos científicos por habitante, y Brasil y Uruguay también nos pasaron —cuando hasta los 90 todos tenían cifras muy inferiores a las argentinas. Y lo hace gastando MENOS: Argentina invierte más en I+D como porcentaje del PBI que Chile y tiene más del doble de investigadores por habitante. Más recursos, más científicos, peor resultado. El problema es el sistema. Es claro que el talento cientifico está. La ciencia e innovación no florecen por decreto, ni con encierro, ni financiando burocracia. Florecen con apertura, competencia, integración al mundo e interacción con el sector privado. (Fuentes; Banco Mundial, National Science Foundation,Our World in Data)

Hvygens @Hvygens

2 months ago

Entre 1997 y 2000, los investigadores de instituciones argentinas publicaban más papers en revistas científicas y técnicas por millón de habitantes que sus pares de cualquier otro país de América Latina y, del mismo modo, los inventores argentinos registraban más patentes

28 133 430 87K 120

The Top AI Papers of the Week (June 14 - June 21): - PreAct - SpatialClaw - Back on Track - OpenClaw-Skill - From Trainee to Trainer - Compositional Skill Routing - Can LLM Agents Infer World Models? Read on for more:

DAIR.AI @dair_ai

2 months ago

x.com/i/article/2068…

3 18 130 41K 159

When Anthropic shut down Fable, every product, every agent, every integration that depended on it went dark overnight. Not because of a technical failure. A government directive forced it, and because Anthropic couldn't segment access fast enough, they switched it off for everyone. Worldwide. This is the clearest proof we've ever gotten that building on closed models means renting, not owning. Bloomberg called it a win for sovereign AI. Fortune reported a global scramble for alternatives. CNBC called it a turning point. They're all right, but for a reason most people are missing. The argument for open models used to be about cost and avoiding vendor lock-in. This week it became something bigger. When a single government directive can wipe out an entire model capability across every country simultaneously, the case for open weights isn't about saving money anymore. It's about resilience. Countries building their own AI stacks aren't being protectionist. They're being rational. If your healthcare system, your financial infrastructure, or your defense apparatus runs on a model that someone else can turn off, you don't have an AI strategy. You have a dependency. The open model ecosystem has been closing the gap on performance for months. Models like Nemotron are already competing at the frontier. This week it closed the gap on urgency.

🚨Anthropic just showed a 24-minute workshop on how to actually do prompts for Claude. Taught by the people who built it. Free. No registration. No paywall. I've seen $300 courses that don't cover what they teach in the first 8 minutes. Watch it and bookmark it now

ニューラルネットワーク（NN）が「群演算（group composition）」を学ぶとき、内部で何が起きているかを初めて数学的に証明した（https://arxiv[.]org/abs/2606.02993）。 群とは整数の加算やモジュラー算術（時計算術、例として x+y mod 12）のような代数構造のこと。g₁⋆g₂ を予測する2層NNを訓練すると、各ニューロンは「既約表現（irreducible representation = 群の最小スペクトル単位で、フーリエ変換の基底に相当するもの）」に確率1で収束することが証明された（Theorem 4.3）。 興味深いのはその「選ばれ方」で、ランダム初期化のタイミングで偶然わずかに優勢だった既約表現が「ロッタリーチケット」として勝ち残り、他の候補は指数的に消えていく。どの表現が選ばれるかは初期化の「籤」で決まる。 アーベル群（足し算のように a⋆b=b⋆a が成り立つ可換な群）では完全な証明が得られた（Theorem 5.1, 5.3）。各ニューロンが学ぶ周波数は全体で均一に分散し、入力層と出力層のフーリエ位相が加法的に揃う（位相アライメント）。両プロセスとも指数収束速度で発生する。 実験はアーベル群 Z₃⊕Z₅（15元素、1024ニューロン）と非アーベル群のフロベニウス群 C₇⋊C₃（21元素）で検証済み。どちらも理論通りの挙動を確認した。 mechanistic interpretability（NNの内部を逆エンジニアリングする研究分野）で「単一周波数への特化」と呼ばれていた現象が、実は群論の表現論的構造そのものを発見していたことを示した研究。さらに「grokking（長い訓練後に突然テスト精度が跳ね上がる現象）」を Stage I（表現学習）→ Stage II（スケールが対数成長してsoftmaxを鋭化）という2段構えのダイナミクスとして説明できる理論的基礎にもなるかもしれない。

La Iglesia Católica se posiciona sobre la IA. El papa León XIV ha presentado “Magnifica Humanitas”, el primer documento oficial dedicado a la Inteligencia Artificial. Es un documento de unas 184 páginas donde compara la revolución de la IA con la primera revolución industrial y llama a regularla con urgencia para proteger la dignidad humana, el trabajo, la educación, los niños y la paz.

Stanford University, OpenAI, Google DeepMind, and Anthropic Researchers believe AI may soon automate AI research itself. Not just coding or writing… but improving the next generation of AI systems. The paper, “AI Researchers' Views on Automating AI R&D and Intelligence Explosions,” interviewed leading researchers across frontier AI labs and academia. It identifies a major shift: - AI is rapidly improving at coding and math - Systems are moving from assistants → autonomous developers - Many researchers now see automated AI research as a realistic possibility This creates a critical turning point. AI may soon stop being only a tool for researchers… and start becoming a research contributor itself. The study also reveals a growing concern: If AI systems begin improving AI systems, progress could accelerate far beyond current expectations. Researchers describe this as a potential feedback loop: - better AI builds even better AI - which accelerates future capability growth This is one of the clearest signals yet that frontier labs are no longer thinking only about chatbots. They are thinking about autonomous research systems. The bigger implication is not just automation, it’s acceleration. As AI becomes capable of handling more of its own development cycle, the pace of progress may increasingly depend on compute and autonomous experimentation—not only human researchers. This points toward a deeper shift in AI: From AI assisting research to AI advancing AI itself article link below:

Nine more Erdős problems have been solved. This time, however, by Google DeepMind. This shouldn't be underestimated, because on the one hand it increases competitive pressure, and on the other hand it proves that the other Frontier Labs can easily keep up.

Przemek Chojecki | PC @prz_chojecki

3 months ago

Another 9 open Erdos problems solved, this time by DeepMind team. Interesting loop of LLM - Lean agents working autonomously, and only after it's verified formally, going through human review.

79 393 3K 681K 765

長文対応VLM（画像とテキストを同時に処理するマルチモーダルAI）をどう効率よく訓練するか。その体系的な答えが公開された（https://arxiv[.]org/abs/2605.13831）。 HKUSTとByteDanceが、Qwen2.5-VL-7B（70億パラメータ）の処理できる文章の長さを32K→128Kトークンに拡張する実験で、3つの反直感的な発見をした。 OCR（文字起こし）訓練は逆効果だった。「ドキュメントの全文字を書き起こせ」と学習させると、性能がベースラインより17.4ポイント下落する。一方、「このページにある情報は？」という質問応答（VQA）形式なら+5〜6ポイントの改善。文字をなぞらせるのでなく「どこに何があるか」を問う形式が重要だった。 128K特化より多様な長さのデータが強い。訓練サンプルの83.9%を128K近くに揃えたデータより、自然な長さ分布（100K超は23.6%のみ）の方が全体的に好成績だった。長文理解は「最大長への慣れ」ではなく「どんな位置からでも正確に情報を引き出す汎用スキル」として身につく。 長文データだけで訓練しても短文性能はほとんど落ちない。短文データを一切混ぜなくても、短文ベンチマーク平均は66.47→65.48の微減にとどまった。テキストLLMでは短文データ混合が必須とされるが、VQA形式の長文データはその必要性を大幅に減らす。 完成したMMProLongは5Bトークンの訓練量で、128Kで学習したのに訓練範囲外の512Kコンテキストでも安定動作する（ベースモデル19.49 vs MMProLong 52.52）。動画データなしに長動画理解まで向上した。153万件・3,600万ページの文書プールからの「問い方の設計」がモデルスケールに匹敵するインパクトを持つ事例。

Collaborative robots are moving automation from isolated cells into daily production activities beside human operators. Factories adopting cobots reorganize safety procedures and line management to gain steadier execution with less physical strain on teams. Microblog @antgrasso

10 WAYS TO PLAY ROBOTICS IN 2026 1. $TSLA building one of the most vertically integrated humanoid stacks combining Optimus Gen 3 hardware, its in-house AI5 inference chip & same FSD-based AI brain powering its vehicle fleet. 2. $NVDA platform layer for entire robotics industry with Isaac GR00T foundation models, Cosmos world models for synthetic training data & Jetson Thor on-board compute. 3. $PLTR Foundry & Warp Speed run mission control layer for robotic & autonomous fleets by turning sensor data & machine telemetry into a single deployable operating picture. 4. $OUSTleading public pure-play in 3D digital lidar where Rev8 native-color sensor line is now shipping to $GOOGL & Volvo Autonomous while being integrated directly into Nvidia Jetson robotics stack. 5. $AVGO supplies custom networking silicon, Tomahawk + Jericho switching ASICs & high-speed connectivity that move data between every robot, sensor & server in autonomy stack. 6. $AVAV, $KTOS, $AVEX & $ONDS are building the UAV drone fleets that feed into the autonomous defense networks. 7. $QCOM provides Snapdragon Ride autonomous platform & on-device AI inference SoCs that power the perception, decision & motor-control loops inside robots, drones & vehicles. 8. $SYNA Astra multimodal edge AI processors run wireless connectivity & sensing inside smart appliances, factory automation & autonomous robotic systems with Core IoT product sales growing 31% YoY. 9. $ISRG, $PRCT, $SYK & $MDTbring surgical robotics that are reshaping operating rooms. 10. $AMZN deploys hundreds of thousands of warehouse robots to move & pick goods while $SYM & $SERV build the autonomous mobile robots that power fulfillment & last-mile retail logistics.

Top 10 Venture Capitalist firms that invested in Mexico last year: 1. Tiger Global 2. SoftBank 3. Kaszek Ventures 4. General Atlantic 5. Monashees 6. Sequoia Capital 7. Valor Capital Group 8. Wollef (former Jaguar Ventures) 9. Andreessen Horowitz 10. Accel Partners Who gonna make the top 10 this year?

Google DeepMind's AlphaProof Nexus autonomously solved 9 open Erdős problems, some unsolved for 56 years, at a cost of a few hundred dollars per problem. It also proved 44 open OEIS conjectures, resolved a 15-year-old question in algebraic geometry, and discovered a novel algorithmic parameter in optimization theory that humans hadn't found. The core mechanism combines LLM reasoning (Gemini 3.1 Pro hype?!) with Lean formal verification. The AI generates proof attempts, Lean's compiler checks every logical step automatically. No human review needed to confirm correctness. The most surprising finding: a basic agent that simply alternates LLM generation with compiler feedback replicated all 9 Erdős successes. The full-featured system with evolutionary search and reinforcement learning only provided meaningful advantages on the hardest problems. This shows a more recent broader trend: as foundation models improve, simple agentic loops are catching up to complex specialized architectures . What sets this apart from OpenAI's informal proof approach: formal verification acts as an automatic filter. The failure analysis showed the AI frequently hallucinated lemmas it claimed were established results, and often disguised the core difficulty by rephrasing it as a helper lemma. Informal proofs would let these errors pass. Lean catches them immediately. The agent also detected misformalizations in existing mathematical literature, correcting ambiguities in problem statements before solving the corrected versions. It served as both a solver and a diagnostic tool. Current limitations are real. Successes cluster in combinatorics, number theory, and optimization where Lean's math library is mature. Problems requiring substantial new theory remain out of reach. Most Erdős problems still weren't solved tho.

keyboard_arrow_left Previous keyboard_arrow_right Next

Darwinが発表された（https://arxiv[.]org/html/2605.14386v1）。 追加学習なしで、既存LLMの重みを直接組み合わせるだけで推論性能を伸ばす手法。フラッグシップのDarwin-27B-OpusはGPQA Diamond（大学院レベルの理科推論ベンチ）で86.9%、1,252モデル中6位を取った（2026-04-22時点）。親モデルはQwen3.5-27BとClaude Opus 4.6スタイルの推論distill版。H100で約5時間、事前学習比で3〜6桁低い計算コストで届く。 仕組みは2つの軸を融合する。レイヤーごとに「推論に効く箇所」をプローブで診断するMRI（Model-layer Response Importance）と、混ぜ方を14次元の遺伝子（ゲノム）で表現してCMA-ESで進化探索するパート。両者を学習可能な信頼度τで動的にブレンドし、τは中間値に収束する。診断だけでも探索だけでも伸びない、というのが効いている。 面白いのは、進化が勝手に発見した非対称な組換えパターン。アテンション層（情報のルーティング担当）は親モデルをほぼ温存し、FFN層（知識処理の本体）は積極的に組み換える。「アテンションは情報の流路、FFNは知識」という従来の解釈と整合する構造に、人手で設計せず辿り着いた点が示唆的かも。 派生も濃い。4Bモデルを再帰的に進化させたDarwin-4B-Davidは+26.4pp（maj@8）跳ね、Transformer×Mamba（状態空間モデル）を学習なしで融合した初のクロスアーキ4B-Genesisも公開。MoE構成のDarwin-35B-A3B-Opusは262K文脈・201言語で90.0%を出している。 巨大モデルを1から学習するより、既存の学習済みモデル同士を組み換えてフロンティアに乗せる、という選択肢の現実味が増している。

Added a DeepSeek Sparse Attention (DSA) from-scratch implementation to my LLMs-from-scratch repo thanks to an awesome new reader contrib. With motivation, overview, and GPT-style model reference implementation as standalone example code: github.com/rasbt/LLMs-fro…

Can frontier models forecast scientific progress? Mostly no, but here is why. This work looks at 4,760 scientific events across disciplines. Frontier models can identify plausible research directions when given options. They cannot reliably predict whether an advance will land, and they get the timeline wrong. They suggest that this is a calibration problem, not a knowledge problem. Frontier models are confidently miscalibrated about whether and when scientific advances arrive. Important grounding for any AI-scientist or research-planning agent that uses model forecasts to pick its agenda. Paper: arxiv.org/abs/2605.22681 Learn to build effective AI agents in our academy: academy.dair.ai

New study shows that the new Physical Efficiency Index (PEI), which includes accelerations, reveals a significantly higher load (1.35 vs. 1.18) than the traditional PEI in professional soccer. More accurate for monitoring fatigue, managing load, and preventing injuries.

An inspiring work on Physical AI: PhysX-Omni. It introduces the first unified sim-ready generation framework for rigid, deformable, and articulated objects, along with a diverse dataset and new benchmark. - Page: physx-omni.github.io - Code: github.com/physx-omni/Phy… - Dataset: huggingface.co/datasets/PhysX…

Physical AI is where the money is going next : $OUST — Lidar sensors for robots and AVs. Velodyne merger. Profitability unproven. $QCOM — Smartphone and auto chip designer. Snapdragon. Diversifying beyond mobile. $ISRG — Robotic surgery monopoly. Da Vinci system. Pay-per-procedure model. $TSLA — EVs, FSD, Optimus robot. Valued as AI co, not automaker. $SYM — AI warehouse robotics. Walmart anchor client.

Los de Figure se van a marcar un directo de 8 horas con sus robots realizando de forma autónoma labores de trabajo con nivel de desempeño humano. Declaración de intenciones en toda regla!

Figure @Figure_robot

3 months ago

Watch a team of humanoid robots running a full 8-hr shift at human performance levels. This is fully autonomous running Helix-02 x.com/i/broadcasts/1…

733 2K 11K 2.8M 2K