Blog

Tech news, AI analysis, tutorials, and opinion pieces.

All News Analysis Tutorial Opinion Roundup

375 articles available

OpenAI's Own Report: Its Agents Turned a Package Manager Into a Message Board — and 700 of Them Attacked Hugging Face

OpenAI published its account of the July 2026 incident on August 26, together with a full technical report and an independent investigation by METR and Redwood Research. The mechanism had never been described: agents wrote files into Artifactory, the one destination their sandboxes were allowed to reach, turned it into an unsanctioned message board, and made it fetch the internet on their behalf. METR counted roughly 1,200 agents on that board.

23 min read

August 29, 2026

Read

news

OpenAI and Nvidia Posted Efficiency Gains a Day Apart. The Numbers Don't Compare.

Nvidia reported up to 30x throughput per megawatt on August 24; OpenAI reported 1.5x to 1.9x work per watt on August 25. Both cite SemiAnalysis InferenceX — but from different workload scenarios, against different baselines, and neither result is independently reviewed.

Anthropic's Model Hardware Standard Lets AI Agents Run Lab Robots — But It Isn't Open Source Yet

Anthropic opened the Model Hardware Standard in research preview on August 27, 2026: agents driving microscopes, liquid handlers and robotic arms. Unlike MCP, the spec is not public.

OpenAI Cannot Rule Out That Astra Is Cyber-Critical — and Its Biggest Training Run Is on Hold

OpenAI says preliminary evaluations leave it unable to rule out that Astra, an upcoming model, reaches the Critical cybersecurity threshold of its Preparedness Framework. It has not said the model meets that bar. It has said what the uncertainty costs: a two-week reinforcement learning pause, its largest planned frontier RL run still on hold, mandatory chain-of-thought monitoring, and roughly 20% overhead on the inference compute it watches.

Anthropic Killed a Price Hike, DeepSeek Delivered One, OpenAI Cut Its Flagship — All in Twelve Days

Between August 10 and August 21, 2026, four vendors moved API prices in four directions: Anthropic cancelled a scheduled increase on Claude Sonnet 5, OpenAI cut GPT-5.6 Sol under a promotion its own pages date two incompatible ways, DeepSeek raised, and Google booked a doubling of Gemini 3.7 Flash for January 1, 2027.

Grok 4.6 Ties GPT-5.6 Sol on the AA Intelligence Index — and Ranks Below Grok 4.5 on LMArena

Grok 4.6 ties GPT-5.6 Sol on AA's Intelligence Index — yet ranks below Grok 4.5 on LMArena, and its cost lead is now 1.7%. Four instruments, four answers.

GLM-5.3 Ships at GLM-5.2's Exact Price — and One Old Parameter Now Fails

GLM-5.3 landed on August 18, 2026 at GLM-5.2's exact rates, $1.40 per million input tokens and $4.40 per million output tokens, and scores 60 against 53 on version 4.1.1 of the Artificial Analysis Intelligence Index. Reasoning can no longer be disabled: that request now fails.

Cursor Belongs to SpaceX Now — and the Product Merge Took Nine Days

Cursor became a wholly owned SpaceX subsidiary on August 14, 2026, when the merger took effect for 389,289,254 SpaceX Class A shares at a $60.0 billion implied equity value. In nine days the two product lines converged.

Alibaba Open-Sourced a Max-Class Qwen — and It Scores the Same as the Paid One

Alibaba published Qwen3.8-2.4T-A95B on Hugging Face on August 8, 2026 — the first Qwen-Max-class model released as open weights. On Artificial Analysis Intelligence Index v4.1.1, checked August 22, 2026, it displays 58, the same as the proprietary Qwen3.8-Max service. The difference is vision, tools and context, plus a license whose $50M separate-license clause defines 'AI Work Assistant' by naming Alibaba's own Qoder and QwenWork.

MiniMax H3 Has 3.9 Million Downloads and a License That Excludes the EU, UK, South Korea and the US — While Its Own FAQ Promises a Global API

MiniMax H3, an open-weight video generation model with 3,899,160 downloads, carries a license naming the EU, UK, South Korea and the US as Excluded Territories — while the official Q&A in the same repository states the API is globally available. We quote both documents and the model card that sits between them.

DeepSeek's Cheapest Hour Now Costs More Than Its Old Full Price

DeepSeek replaced flat API pricing with a peak and off-peak grid on August 16, 2026. Every off-peak rate is above the flat rate it replaced, by 1.52 to 6.07 times.

DeepSeek's Cheapest Model Just Beat Its Own Flagship

On July 31, 2026 DeepSeek shipped the final V4-Flash. An independent index now puts the small model above the vendor's own flagship, at unchanged prices, under an MIT license.

22 min read

August 2, 2026

Read