Live Feed
OpenAI Is Retiring the Entire GPT-4.1 Era: The 2026 Model Graveyard
OpenAI's official deprecation log confirms the full GPT-4.1 family is being wound down in 2026. GPT-4.1-nano fine-tuning windows close in roughly three months. Only GPT-5.x models survive. Developers on gpt-4.1-mini or gpt-4.1 need a migration plan.
GPT-5.6 Sol Pro Closes a 30-Year Gap in Convex Optimization, Lean-Verified in 2.5 Hours
UC Berkeley professor Phillip Kerger used GPT-5.6 Sol Pro to prove that d² function evaluations are the minimum needed for zeroth-order convex optimization, closing a gap open since Protasov's 1996 upper bound. The proof checked out in Lean.
Kimi K3 Takes Frontend Code Arena #1 at 1679 — First Chinese Model to Beat Fable 5 and GPT-5.6
Moonshot AI's 2.8-trillion-parameter K3 jumped 17 places to top Arena's Frontend Code leaderboard with 1679 ELO, beating Claude Fable 5 and GPT-5.6 Sol in blind developer testing. Full weights ship July 27.
Altman Warns GPT-5.6 Sol May Hit Infrastructure Hiccups as Inference Demand Outruns Capacity
OpenAI's CEO posted on X that Sol demand is 'insane' and warned that 'hiccups' are possible as the inference team struggles to scale. The warning came four days after Sol's public launch and surfaces the gap between training a frontier model and running it at demand-matching scale.
StackOverflow's March 2026 Question Count Is 29x Below Its 2017 Peak — and Below Its Own Launch Month
Public Stack Exchange data shows the monthly question rate has collapsed from 286,300 in March 2017 to 9,883 in March 2026. May 2026 answer volume fell below StackOverflow's June 2008 launch baseline. The platform that trained every AI coding model is being replaced by the tools it enabled.
New York Freezes Hyperscale Data Center Permits — Hochul's July 14 Order Is the First US Statewide AI Infrastructure Moratorium
Governor Hochul signed an executive order on July 14 suspending permits for any new data center drawing more than 50 megawatts, while regulators write binding standards on energy, water, and environmental impact. The yearlong pause is the first statewide moratorium on AI infrastructure in the United States.
Pure DC Breaks Ground on €7.5B Finland AI Campus — Phase 1 Fully Leased at 110MW
UK-headquartered Pure DC is building one of Finland's largest-ever inward investment projects: a 550MW+ AI campus in Seinäjoki with €7.5B total potential. Phase 1, at 110MW and €1.5B, is fully contracted and already has its substation live.
Arena Starts Scoring for Truth: Factuality Now Weighted Alongside Human Preference
Chatbot Arena is adding factuality as a ranked signal in its Text and Search leaderboards, using AI-verified claim accuracy to supplement human preference votes. Models that win on vibes but hallucinate may drop in the unified ranking.
Washington Eyes Open-Source AI Capability Cap Tied to China's Best Models — Beijing Calls for More Openness
The Trump administration and the AI industry are discussing a capability framework for US open-source models based on current Chinese open-source capabilities. Industry expects Chinese Mythos-class weights to eventually be downloadable. At the World AI Conference in Shanghai, Xi Jinping said China is ready to be more open about AI.
Kimi K3's First Benchmark Numbers: 90.7 LiveBench Reasoning, #4 on AA Intelligence Index at $0.38 Per Task
Moonshot AI's open-weight 2.8T flagship posts first independent benchmark results: 90.7 reasoning on LiveBench — 1 point behind GPT-5.6 Sol Max Effort and above Fable 5 — and lands 4th on Artificial Analysis Intelligence Index. At $0.379 per successful task, it is the cheapest model in the LiveBench top six.
Apple Sends Legal Letters to Dozens of OpenAI Employees in Trade Secret Escalation
Weeks after suing OpenAI and naming two senior employees, Apple has sent legal letters to dozens more current OpenAI workers it alleges left Cupertino carrying confidential information. The action broadens the IP dispute from a corporate lawsuit into an individual-employee campaign.
GPT-5.6 Is Burning Through $200/Month Codex Pro Accounts in Hours — OpenAI Has No Fix
Pro subscribers report GPT-5.6 consumes dramatically more tokens for equivalent tasks than GPT-5.5, with some losing their full 5-hour weekly allowance after two tasks and over $100 in additional purchased credits in a single session. OpenAI support calls it normal.
AI Absorbed $222B and 66% of US Venture Capital in 2025 — Up From 10% a Decade Ago
PitchBook's 2025 AI investment landscape puts a precise number on the structural shift: AI captured 65.6% of all US VC deal value last year, up from 47.2% in 2024 and 10% in 2015. US AI unicorns now aggregate to $4.3 trillion in value.
NVIDIA Nemotron-3-Embed Tops RTEB at 78.5%: Three Models to Fix Agent Retrieval
NVIDIA ships three open embedding models targeting the retrieval bottleneck in agentic AI. The 8B flagship takes RTEB #1 at 78.5%; two 1B variants cover cost and latency tiers with a Blackwell-optimized build hitting 2x throughput at 99%+ of BF16 accuracy.
LMCache Cuts Agentic Inference Latency 3x — KV Cache Layer Addresses Coding Agent's Prefix Recompute Problem
LMCache is a tiered KV cache management layer for vLLM that eliminates repeated prefix recomputation in agentic workloads. At 32 concurrent users on 100K-token coding agent traces, it delivers 3x lower average latency and 2.3x more completed requests versus GPU-only prefix caching.