AI & ML

Do DeepSeek and Qwen need different prompts from ChatGPT

"Avoid adding a system prompt" is the most-repeated rule about prompting Chinese models, and DeepSeek's own current documentation opens its examples with one. We read what all eight major Chinese vendors publish: not one gives a rule about phrasing a prompt in Chinese, and seven of the eight now think by default, which retires most of the knobs the advice is about.

27 Sep 2026 · 8 min read

AI & ML

Anthropic Cut Opus Prices for the First Time, and OpenAI Halved Sol Hours Later

Opus 5.5 lands at US$4 and US$20 per million tokens, a fifth below the model it replaces. GPT-6 Sol arrived that morning at exactly half those rates, which OpenAI calls permanent.

26 Sep 2026 · 10 min read

AI & ML

xAI's Grok 4.7 Leads Two of Seven Benchmarks on Its Own Scorecard

Fable 5.1 takes four and GPT-5.6 Sol one. The wins are electrical engineering and a legal agent test, where the margin is threefold.

25 Sep 2026 · 4 min read

AI & ML

How many tokens is a Chinese character, and what Chinese AI companies call things

Four Chinese model vendors publish the same ratio in their own documentation: a Chinese character is about 0.6 of a token, not one. We then counted which Chinese words eight vendors actually write across 730,128 characters of their docs — 词元, the approved word for token, appears zero times, and 令牌 usually means an access credential.

25 Sep 2026 · 8 min read

AI & ML

Codex Read a Repository and Ran Its Owner's Code

Two escapes from OpenAI's coding agent were reported in August and fixed within eight days. The more serious one worked from the mode that grants no write access at all.

22 Sep 2026 · 5 min read

AI & ML

Canada and Germany Put C$300m Behind an AI Built to Watch Other AIs

LawZero's Scientist AI is designed to have no goals of its own. Two states are funding oversight because they cannot outspend the labs on capability.

17 Sep 2026 · 5 min read

AI & ML

Shanghai AI Lab Quietly Released a 744B Agent Model Under an MIT Licence

No blog post, no pricing, and a paper claiming the best score on five of 16 benchmarks without saying which five. The human study is the part worth reading.

16 Sep 2026 · 6 min read

AI & ML

California Will Make Companion Chatbots Pass Independent Child-Safety Audits

Thirteen bills signed on 10 September also ban history-based feeds for under-16s. The audit rule is the new part, and California has not yet said who the auditors will be.

15 Sep 2026 · 6 min read

AI & ML

Salesforce Named Its Agents. One of Them Works a Customer for Weeks.

Six of the seven are generally available and Hunter follows in November. Its long-horizon runtime pursues a goal over weeks — a different product from the other six.

14 Sep 2026 · 7 min read

AI & ML

Meta Bought Stilla.ai, an Eight-Month-Old Startup, to Run WhatsApp's Shop Counter

Stilla.ai left stealth in January on $5m. The price is undisclosed, the durable asset is two ex-Shopify founders, and the volume that matters is in Southeast Asia.

11 Sep 2026 · 5 min read