Live Feed
Celeris-1 Hits 1,664 Tokens/Second: A Diffusion LLM Enters the Production API Tier
The second commercially available diffusion language model posts 158ms p50 response times at 75.9% MMLU-Pro accuracy — 5x Mercury 2 on speed, 12 points ahead on quality.
Fubon: Google Plans 12–15 Million TPU v9 Chips by 2028, Intel Foundry Required
The volume would approach Nvidia's projected 12.4 million GPU shipments in the same year. TPU v9 shifts to four compute dies per package, more than doubling capacity consumption from 2027 and forcing Google to split production between TSMC and Intel.
GPT-5.6 Sol Triples on ARC-AGI-3 Without a Weight Change — Two API Settings Did It
Retained reasoning and compaction moved Sol from 13.3% to 38.3% on the public task set while cutting output tokens sixfold. On the official ARC Prize leaderboard, Sol is at 7.8% and Opus 5 at 30.2%.
DeepSeek Plans 1GW Data Center in Inner Mongolia — Efficiency Lab Bets on Infrastructure
DeepSeek is building at least 1GW of AI compute capacity in Ulanqab, Inner Mongolia, combining its own facility with leased capacity from third-party operators. Partial operations by late 2027. The company has raised $7.4B from Tencent, CATL, and others to fund the shift.
Meta: Naive RL Fails for Code Optimization. Rebuilding the Timing Pipeline Lifted Qwen 2.5 7B by 13 Points.
A new Meta paper diagnoses why standard RL for code speed optimization almost always fails: runtime is a noisy, sparse signal, and small defects anywhere in the evaluation chain corrupt the training. A full pipeline rebuild took Qwen 2.5 7B from 18.0% to 31.3% on the top-50% speed threshold.
Nvidia Forms Open Secure AI Alliance: GLM 5.2 Cleaned Up What Fable 5 Refused to Touch
Nvidia led dozens of AI companies in forming OSAA to build open-source defensive cybersecurity AI. The immediate trigger: Fable 5 refused to assist with HuggingFace incident response, and Nvidia's quantized build of GLM 5.2 completed the task instead.
AMD Trains Its Own MoE on Instinct GPUs and Opens Every Checkpoint — But the Weights Aren't Commercially Free
AMD released Instella-MoE-16B-A3B, a 16B MoE with 2.8B active parameters trained end-to-end on MI300X and MI325X with ROCm. Every training stage, data mixture, and config is published. ResearchRAIL limits commercial weight use; the MIT-licensed training code is fully open.
South Korea's Two 700B+ Models Release 48 Hours Apart — Elimination Round Starts August 8
SK Telecom's A.X K2 (688B) and LG AI Research's K-EXAONE 2.0 (750B) both shipped under Apache 2.0 within 48 hours. A.X K2 leads on Korean and math; K-EXAONE 2.0 leads on long-context retrieval. A.X K2 scores 9.3 on BrowseComp. Government evaluation August 8-11 cuts the field from four to three.
Disney Drops GitHub Copilot for OpenAI Codex as Enterprise Coding Market Consolidates
Disney is pulling GitHub Copilot and several other AI coding tools from US engineering teams in August, standardizing on OpenAI Codex. The same week, Cursor removed cost transparency from its usage page ahead of an expected SpaceX acquisition close.
Explorative Models Train on the Best of K Guesses — A New Pretraining Axis Targets End-to-End Generation
A new pretraining paradigm generates K candidate outputs at each training step and learns only from the best match, eliminating mode blurring without requiring VAE or codec pipelines. Researchers call exploration a third pretraining axis alongside scale and architecture.
Nvidia in Talks to Finance $250B of a 10GW OpenAI Data Center in Ohio
Nvidia is in advanced discussions to guarantee $250 billion in financing for a SoftBank-developed 10-gigawatt AI data center in southern Ohio, the largest compute project ever announced. OpenAI is the primary tenant in talks; Anthropic, Microsoft, and Google have also contacted Commerce Secretary Howard Lutnick about the project.
OpenAI Names Astra as Its Next Major Model — Pre-Release Solves 10 Unsolved Math Problems
An internal build of Astra, OpenAI's next model family, resolved ten open problems spanning geometry, coding theory, group theory, and post-quantum cryptography. Each problem had been open for at least a decade. All proofs are Lean-formalized and required roughly $2,000 in compute at Sol API rates.
Morgan Stanley Leads $15B Debt Package for Anthropic's Texas Campus — Google Takes 20% Equity in the Project
A banking group led by Morgan Stanley is in talks to provide a $15 billion financing package for a Nexus Data Centers campus in Hubbard, Texas tied to Anthropic. Google will guarantee lease and power obligations in exchange for roughly 20% equity in the data center project itself.
Claude Opus 5 Tops LiveBench Agentic Coding at 65.2% — Fable 5 Trails by 3 Points at Twice the Cost
On LiveBench's agentic coding dimension, Claude Opus 5 Thinking Max Effort scores 65.2%, above Fable 5 at 62.2%, Kimi K3 at 62.2%, and GPT-5.6 Sol at 56.2%. The gap matters: Opus 5 costs $0.699 per successful task while Fable 5 costs $1.439.
AMD Locks 529MW From Core Scientific in $14B Lease — Infrastructure Is the New AI Moat
AMD has secured 529 megawatts of US AI data center capacity from Core Scientific under 15-year leases worth more than $14 billion. The deal is AMD's largest infrastructure commitment and signals a strategic shift: chip designers are building the physical layer, not just the silicon.