AI Daily Brief — 23 December 2025
Google DeepMind launches Gemini for Government — accelerated access for all 17 DOE National Labs under Genesis Mission; AI co-scientist multi-agent live; AlphaEvolve + AlphaGenome early…
Google DeepMind launches Gemini for Government — accelerated access for all 17 DOE National Labs under Genesis Mission; AI co-scientist multi-agent live; AlphaEvolve + AlphaGenome early…
Z.ai opens GLM-4.7 (~400B, 200K context, SWE-Verified 73.8% matches/beats Sonnet 4.5 on LiveCodeBench v6 + Terminal Bench 2.0). Anthropic ships Bloom — open-source agentic framework for…
Sunday quiet. IBM CUGA on HF Spaces — dynamic task ledger + re-planning + sandboxed pseudo-code + OpenAPI/MCP/LangChain. Child-safety AI companions debate (children using AI 42%…
Tesla Optimus v2.3 Berlin debut at Mall of Berlin (10am-8pm) — second European stop after London Dec 13; first 50 attendees received Tesla-branded Christmas gifts. 23-AG…
OpenAI ships GPT-5.2-Codex — SOTA SWE-Bench Pro + Terminal-Bench 2.0; long-horizon via context compaction. OpenAI FrontierScience benchmark (700+ Q; 42 medalists + 45 PhDs): GPT-5.2 leads…
Anthropic publishes Agent Skills as open standard at agentskills.io (spec + SDK). Skills directory live with Atlassian + Canva + Notion + Figma + Cloudflare +…
Google launches Gemini 3 Flash — default model Gemini app + AI Mode in Search global (replaces 2.5 Flash). SWE-bench Verified 78% beats Gemini 3 Pro;…
OpenAI ships GPT Image 1.5 — faster + more precise editing + 4× speed + 20% cheaper API. OpenAI publishes FrontierScience benchmark: GPT-5.2 leads with 77%…
NVIDIA debuts Nemotron 3 family open models (Nano/Super/Ultra) — hybrid MoE + Mamba-Transformer, 1M context, 4× throughput vs Nemotron 2; 3T tokens pretraining + 18M post-training…
Sunday pure digestion. Christopher Penn '2025 AI Year in Review': Gemini 3 smartest overall; Claude Opus 4.5 smartest coder; Nano Banana / Pro breakout image; PhD-level…
Quiet Saturday after Dec 11 trifecta. Tesla Optimus v2.3 debuts outside US at Westfield London (8am-8pm; popcorn + selfies + 50 Tesla-branded Christmas gifts). Disney-OpenAI $1B…
Dec 11 trifecta ripples: Disney-OpenAI analysis pieces (WGA: 'theft of our work'). Newsom: Trump EO attempt to 'enrich himself + his associates'. NIST opens 45-day comment…
Trifecta day. OpenAI ships GPT-5.2 (Instant/Thinking/Pro; GDPval 70.9% vs 38.8% GPT-5.1; SWE-Bench Pro 55.6% beats Gemini 3 Pro + Opus 4.5; 400K context + 256K MRCR;…
Google launches fully-managed remote MCP servers (Maps + BigQuery + Compute Engine + GKE) — unified governed agent layer. Microsoft AI publishes Copilot Usage Report 2025…
Accenture + Anthropic launch 3-year partnership — 30K Accenture pros trained on Claude + Claude Code center. OpenAI ships GPT-5.2 in 'Code Red' response to Gemini…
OpenAI Sora 2 snapshot bump (sora-2-2025-12-08). NextEra + Google Cloud unveil multi-GW DC partnership (3 sites + TimesFM 2.5/WeatherNext 2 internal). Microsoft enables Anthropic-as-subprocessor M365 admin…
Sunday very quiet: NeurIPS 2025 final day in San Diego. No frontier-lab releases. OpenAI's '12 Days of OpenAI' (Shipmas) Day 3 queued for Mon. Pentagon-Grok backlash…
Quiet Saturday: McDonald's Netherlands releases AI-generated Christmas ad 'It's the Most Terrible Time of the Year' (TBWA + Sweetshop) — immediate backlash, made private Dec 9.…
Meta strikes commercial AI data deals with CNN + Fox + Le Monde + People + USA Today + 3 more — first major Western frontier-lab…
Google ships Gemini 3 Deep Think to Ultra subscribers ($250/mo) — HLE 41%, ARC-AGI-2 45.1%. AWS re:Invent Day 4: 3 frontier agents (Kiro + Security +…
AWS re:Invent Day 3: Swami Sivasubramanian agentic-AI keynote. Snowflake + Anthropic $200M strategic partnership (Claude for 12,600+ enterprises via Cortex). AWS Trainium3 UltraServers GA (3nm, 2.52…
Mistral ships Mistral 3 family open Apache 2.0 — 675B/41B-active MoE flagship + Ministral 3B/8B/14B + 256K context. Anthropic launches Claude for Nonprofits on GivingTuesday (75%…
DeepSeek releases V3.2 + V3.2-Speciale open under MIT (685B, AIME 96.0%, IMO gold 35/42, ICPC 2nd) with DeepSeek Sparse Attention paper (O(kL) near-linear). Cyber Monday all-time…
Adversarial Poetry paper goes viral: 62% jailbreak success across 25 frontier models (Gemini + DeepSeek + Mistral most permissive; GPT-5 + Claude Haiku 4.5 most resistant;…