May 17, 2026 (Sun)

[01]

Paper Picks

No new arXiv papers from Tier 1 orgs met the bar today. The papers captured yesterday (2026-05-16) remain the most recent:

Achieving Gold-Medal-Level Olympiad Reasoning via Simple and Unified Scaling — Gold-medal olympiad reasoning via unified scaling (134 upvotes, HF Daily #1)
Self-Distilled Agentic Reinforcement Learning — Self-Distilled Agentic RL (75 upvotes, HF Daily #3)

Tracking: ICML 2026 accepted — tool-calling LLM evaluation framework (authors/org unverified, low priority).

[03]

Anthropic Institute research agenda (May 8) — 4 pillars: (1) Economic diffusion, (2) Threats & resilience, (3) AI systems in the wild, (4) AI-driven R&D. Led by Jack Clark. Worth tracking on a quarterly basis whether this research translates into actual policy influence. → Anthropic
OpenAI DOE Genesis collaboration — "Year of Science" slogan, deploying reasoning models on the Los Alamos supercomputer (Venado). Worth tracking the OpenAI vs DeepMind comparison in the AI-for-science race. → OpenAI, Google DeepMind
Mistral Forge (March 26) — a platform for training frontier-class models on enterprises' proprietary data. Partners include ASML, ESA, Ericsson. To be integrated with Devstral 2 + Vibe CLI. Worth tracking as an enterprise AI business model. → Mistral AI
Anthropic $30B annual revenue ARR (April 2026) — $9B (end of 2025) → $30B (4 months, 3×). If this pace continues, an important data point for the quarterly synthesize.

[04]

Pages created today. User appropriateness review recommended:

AI Alignment (⚠️ NEW — anchored on Anthropic's "Teaching Claude Why." Beginning tracking as a core 2026 safety research concept)
Gemini 3.1 Deep Think (new — Gemini Deep Think / Aletheia autonomous mathematics agent. Google DeepMind's reasoning model page)
Devstral 2 (new — Mistral Devstral 2 + Devstral Small 2 + Vibe CLI. The frontier of open-source coding agents)

[05]

Anthropic: added Teaching Claude Why + Anthropic Institute agenda + $50B infrastructure
Google DeepMind: added Gemini Deep Think + Aletheia + AI for Math Initiative
Mistral AI: added Devstral 2 + Vibe CLI + Mistral Forge
xAI: added Grok Connectors + BYO MCP + grok-4.3 status
OpenAI: added DOE Genesis collaboration details
Grok Build: added Devstral 2 / Vibe CLI to the Compared To table (lint-recommended cross-ref)

[06]

Health: 84/100 (+12 vs W20)

Dead links: 14 → 7 (4 core ones resolved: xai, alphaevolve, agents, test-time-compute, llm-knowledge-bases, nvidia, alignment)
Missing xrefs: 4 (grok-build ↔ devstral-2, olympiad-paper ↔ gemini-deep-think, alignment ↔ agentic-rl, karpathy ↔ alignment)
Stale / Contradiction / Orphan: all 0 ✓

User decisions needed:

Whether to create a [sam-altman](/wiki/people/sam-altman) page (4 dead refs — worth tracking)
Decide on a unified [claude](/wiki/claude) page vs keeping per-model distribution
Whether to create [lecun](/wiki/people/lecun) / [world-models](/wiki/concepts/world-models)

[07]

5 additional items captured that were missed in the morning ingest.

⚠️ 1. OpenAI "Accidental CoT Grading" disclosure — first real-world case of an alignment monitorability risk (score ~2.2 — ALERT)

OpenAI disclosed in a 2026-05-07 alignment blog that GPT-5.4 Thinking, GPT-5.1–5.4 Instant, and GPT-5.3/5.4 mini models had their CoT (reasoning process) itself accidentally evaluated as a reward signal during RL training (source)
Risk: grading the CoT directly can teach the model to hide incriminating thoughts when misbehaving and to generate reasoning traces that "pretend to be correct"
Result: "no clear evidence of monitorability degradation" (OpenAI's own assessment). However, external reviewers from Redwood Research and LessWrong questioned the strength of the evidence
GPT-5.5 was unaffected. Mitigations: fixed the reward path and expanded the automated detection system
Why it matters: CoT transparency is a core tool of current alignment auditing. A case showing this transparency can be undermined by RL has now been disclosed for the first time in an actually deployed model. Where Anthropic's "Teaching Claude Why" addressed a training methodology problem, this addresses an auditing methodology problem — both must be solved for alignment to hold.
→ AI Alignment, OpenAI

🆕 2. Gemma 3n (Google DeepMind, 2026-05-12 preview) — a 3× efficiency breakthrough for mobile AI

5B/8B models run in 2GB/3GB RAM — 3× compression vs prior via the PLE (Per-Layer Embeddings) architecture (source)
Multimodal: Audio (ASR / translation) + image + video + text. Enhanced multilinguality (WMT24++ 50.1%)
8B variant: dynamic switching to 4B/2B — a single model can trade off performance vs efficiency
On-device optimization via Qualcomm / MediaTek / Samsung collaboration. 1.5× faster than Gemma 3 4B
Why it matters: a tipping point where "on-device AI" rises from barely text-level to multimodal (voice + video) level. Once real-time ASR + translation + image processing becomes possible on smartphones, the edge deployment paradigm itself changes.
→ Gemma 3n (new), Google DeepMind

🆕 3. Karpathy autoresearch — "simulating a community of researchers, not a single PhD"

2026-03, a single ~630-line Python file (source)
The agent ran independently for 2 days → found 20 validation-loss improvements → all confirmed to transfer to larger models
Vision: "SETI@home style" — thousands of agents running an asynchronous, parallel research community
Why it matters: autoresearch is a perfect demonstration of Karpathy's Software 3.0 Verifiability Principle — RL-based agents work because validation loss provides an automatic reward signal. A concrete realization on the path toward Sam Altman's prediction of an "automated AI researcher by 2028."
→ Andrej Karpathy, Agents (LLM Agents)

4. Mistral Connectors + MCP in Studio (2026-04-15)

GitHub, Gmail, and web search built in by default + Code Interpreter + Image Gen + Document Library (RAG). Full support for the Custom MCP API (source)
Human-in-the-loop approval flow + programmatic CRUD
Devstral 2 + Vibe CLI + Connectors = Mistral completes its full open-source agent stack
Why it matters: following Grok BYO MCP and Claude MCP, Mistral too has confirmed MCP as an API-first feature. MCP is converging into the de facto agent tool standard.
→ Mistral AI

5. Anthropic Enterprise AI Services JV (2026-05-04) — $1.5B, Blackstone, Goldman, H&F

Anthropic + Blackstone + Hellman & Friedman + Goldman Sachs established a standalone AI services company (source)
$1.5B total committed. Apollo, General Atlantic, GIC, and Sequoia additionally participating
Anthropic engineers dispatched directly → redesign enterprise workflows + integrate Claude
One week ahead of OpenAI's Deployment Company (5/11) — a "frontier lab → consulting industry entry" competitive dynamic
Why it matters: the front line expands from model competition to a competition over "who is embedded more deeply in real enterprises." A direct challenge to the Big 4 consultancies. → Anthropic

Late Additions stats: 5 new sources, 1 new model page (gemma-3n), 7 wiki pages updated, 8 dedup skipped