LLM Benchmark Watch

Tracking Kimi's coding model push:

• K2.6 announced: New model on the way, tied to KIMI Code Beta.
• Preview live: K2.6-code-preview now available as...

Mixed signals on Anthropic's Claude: reliability crumbling amid outages and escalating quality complaints, yet Mythos Preview hailed as too powerful...

Real-world AI adoption in gov procurement: GSA, DoD, IRS (AICDT for ML doc review), and Army (DORA, now required by AFARS) use AI to scan forms,...

Unified multimodal models exhibit pseudo-unification, as entropy probing reveals divergent information patterns. A key eval limitation for foundation model architects.

Key takeaways from Stanford's 2026 AI Index:

• No performance plateau: Models show accelerating leaps on benchmarks, debunking slowdown claims.
  -...

Benchmark hack exposed:

• UC Berkeley researchers built an AI agent that hacked eight major AI benchmarks for near-perfect scores
• Includes SWE-bench Pro
• Cheating undermines eval integrity for LLM leaderboards

• Conversation reset & mgmt: /CLEAR resets history, /COMPACT compresses chats, /BRANCH creates branches, /REWIND rolls back.
• Project tooling: /INIT...

Apple counters AI investment mania with strategic restraint and a war chest:

• Financial firepower: $400B revenue, $100B FCF, $65B cash + $77B...

• Qwopus drops: Mimics Claude Opus 4.6 reasoning on regular laptops, no cloud/hardware costs—ideal for devs/hobbyists.
• Darwin-27B-Opus...

Enterprise agent boost: Testing OpenClaw-like features in M365 Copilot with superior security for businesses.

• Autonomous power: Always-on agent for...

New paper introduces Process Reward Agents for steering knowledge-intensive reasoning in LLMs, targeting complex reasoning chains. Links: https://t.co/3JPG5C99Xx https://t.co/dRCKq3AOkM.

Milestone achievement: Claude Mythos Preview is the first model to fully complete "The Last Ones" (TLO), a 32-step corporate network attack sim...

• Major login outage struck Claude ~11:45-12:15 ET, with Downdetector peaking at 4,000+ reports before dropping sharply.
• Anthropic responded...

Enterprise AI is shifting from model wars to system-level reliability for complex workflows.

• Simulation pillar: eVerse stress-tests agents on edge...

Key perception gaps from Stanford's AI report signal adoption challenges:

• Overall US impact: 56% experts positive vs only 10% public more excited...

Mistral Forge empowers enterprises with complete training pipelines on proprietary data, unlike API services.

Key advantages:

• Domain embedding:...

Key breakthroughs for enterprise AI agent orchestration in BoxLang AI 3.0:

• Agent trees: Full hierarchy with parentAgent, depth queries...

FORGE launches as a fine-grained multimodal evaluation benchmark tailored for manufacturing scenarios, extending LLM assessments beyond text to industrial real-world use.

Anthropic boosts Claude Code for proactive assistance via community Skills and Epitaxy upgrade:

• Rails audit Skill: Open-source...

Benchmark shakeup: Anthropic's top model leads China's best by just 2.7 percentage points as of March 2026, after DeepSeek R1 briefly matched U.S....