Live Feed
Aleph Alpha's Kolibri: 78B-Parameter Sovereign EU LLM Ships Under Apache 2.0
Aleph Alpha released Kolibri 1 on October 3 -- a 78 billion total parameter mixture-of-experts model trained on infrastructure in Germany and Finland, EU AI Act compliant from the ground up, bilingual German-English, with 262K context and Apache 2.0 weights.
OKI Pushed PCB Layer Count to 180 for HBM Testing — Up 45% in 18 Months
OKI Circuit Technology developed 180-layer, 15mm-thick printed circuit boards for HBM wafer-testing equipment in April 2026, extending a 2025 breakthrough with 124-layer boards. As AI accelerators stack more HBM dies, the test hardware qualifying every chip has become its own substrate race.
SemiAnalysis Cut Google's Ironwood TPU Forecast by 500,000 Units. Alphabet Is Now Filling the Gap With Third-Party Compute.
SemiAnalysis slashed its 2026 TPU v7 (Ironwood) production estimate from 3.2 million to 2.7 million units in August, citing CoWoS-S packaging constraints. On its Q2 2026 earnings call, Alphabet confirmed it will expand third-party compute capacity in Q3 — an admission that in-house chips alone cannot cover demand.
Meta Open-Sources Muse Gadgets SDK to Put Its AI Agent on ESP32 and Raspberry Pi
Meta launched Muse Gadgets, an open-source project that lets developers connect the Muse AI agent to custom hardware including ESP32 microcontrollers and Raspberry Pi displays. The move extends Muse beyond Meta's own Muse Home Link smart device and the upcoming Muse Charm wearable.
Ai2 Open-Sources AstaBrief 8B: Cited Scientific Reports in 51 Seconds vs 178 for Claude
The Allen Institute for AI released AstaBrief 8B, an 8-billion-parameter model trained on tens of thousands of real research queries to turn literature excerpts into cited reports in a single pass. Deployed as Fast mode in Asta, it runs 3.5x faster than the Claude-powered Thinking mode and can run on local infrastructure for sensitive work.
OpenRouter's Router Benchmark Shows 48x Cost Gap Between Cheapest and Priciest Routing Strategies
OpenRouter published its first systematic comparison of model routers, scoring six routing strategies on a composite index weighting quality 60%, speed 20%, and cost 20%. Routing to DeepSeek V4 Flash instead of GPT-6 Astra cuts per-token cost 48x and makes a 10-49 turn Codex session 21x cheaper.
Ataraxos AI Defeats Stratego World Champion 15-1 on 16 GPUs — Cracking 10^66 State Space
A system called Ataraxos beats Stratego world champion Pim Niemeijer 15-1, solving the last major tabletop game where AI had no credible win. The key: a second neural network that infers the identity of hidden opponent pieces, tackling imperfect information at a game-tree scale of 10^66 configurations.
Supabase Raises $150M and Acquires Turso to Provision Per-Agent Databases at Scale
Supabase closes a $150M round led by GIC and acquires Turso, whose architecture provisions on-demand isolated databases per agent at low cost. The deal positions Supabase as the default database layer for agentic AI workloads in both public cloud and BYOC environments.
FLUX 3 Image Ships with Non-Destructive Multi-Step Editing and Bounding-Box Subject Control
Black Forest Labs releases FLUX 3 Image, the dedicated image endpoint of its FLUX 3 multimodal family. Supports multi-step edits that leave surrounding pixels untouched, bounding-box subject positioning, and up to 4K output. Open-source weights expected within weeks.
Harvard Physicist Found the Right Problems for Claude: 15 New Feynman Integrals, 36 Papers in 90 Days
Matthew Schwartz built BootLoops, a Claude harness for exact quantitative science. In three months it produced 30 scattering-amplitude integrals — 15 never computed before — spread across 36 manuscripts in 18 fields.
Upstage Solar Mini 4: Intelligence Index 24 at $0.10 Per Million Tokens, 2x the Median for Its Price Tier
Upstage's Solar Mini 4 scores 24 on Artificial Analysis Intelligence Index, twice the median of 12 for reasoning models at its price tier, at $0.10 input and $0.40 output per million tokens. The model targets high-volume workloads where frontier pricing is not viable.
DeepSeek Harness v0.2 Goes Desktop: macOS and Windows Installers, Plugin Manager, Scheduled Automations
DeepSeek's Harness coding agent ships v0.2 preview with native desktop installers for macOS and Windows, adding a plugin manager and scheduled automations. It is already the most popular coding agent on DeepSeek's official API by daily active users and daily sessions.
Aalo Atomics and Crusoe Commit $200M to Nuclear-Powered AI Factory, Idaho Demo in 2027
Nuclear reactor startup Aalo Atomics and cloud provider Crusoe have committed more than $200 million to build the first nuclear-powered AI data center. A live demonstration is targeted at Idaho National Laboratory in 2027, with commercial 50 MW reactor pods planned for deployment by end of 2029.
arXiv Caps Submissions at 2 Per Month After AI Doubles Volume to 40,363 in September
arXiv received 40,363 submissions in September 2026, double the September 2024 pace, generating 9,000 moderator support tickets. Effective October 1st, the repository limits all submitters to two papers per calendar month and three active submissions at any time.
McDonald's ML Engine Prices Big Macs by Neighborhood Willingness to Pay
McDonald's machine-learning system analyzes millions of transactions across 14,000 US restaurants and estimates what each local market will accept. Two Fresno locations two miles apart already show a $1.20, 21% price gap on the same item.