Oracle bans AI-generated code from OpenJDK

A policy line drawn in Java, a benchmark for tutor restraint, and a free speedup for long-context decoding on Intel's Battlemage GPUs — the day spans governance, pedagogy, and raw performance.

Oracle has told OpenJDK contributors not to submit AI-generated code or content — an interim policy that also covers documentation, pull requests, emails, and bug reports. The company cites review burden, the risk of "plausible-looking but incorrect code" in mission-critical systems, and unresolved copyright questions under the Oracle Contributor Agreement; AI tools remain fine for debugging, analysis, and review, just not for authorship. The awkwardness is hard to miss: Oracle is pouring $70 billion into AI infrastructure while its executives tout AI-written internal code — yet it wants none of it in the Java it stewards. A permanent policy for the OpenJDK Governing Board is in the works, and other foundations

Ai2 released TutorMoments, a replay-based evaluation that tests whether LLM tutors know when to help and when to hold back. Built on 462 de-identified transcripts of real one-on-one math tutoring (US grades 2–7) with 1,500+ teacher-flagged decision points, it pauses a session at a key moment and lets a model tutor a simulated student for five turns, scored on scaffolding, pushing for rigor, and over-scaffolding. The finding: models default to over-helping, and even when the trade-off is spelled out in the prompt, they don't close the gap to human tutors who vary their moves — the dataset deliberately concentrates on moments where tutoring could have gone better, so the low human scores aren't a ceiling. A useful corrective to benchmarks that reward one fixed behavior

Long-context decoding on Intel's Battlemage GPUs just got up to 169% faster — from a single routing change. An open llama.cpp change re-routes how the runtime reads its quantized key-value cache on the new Intel cards, measured at +42% to +169% faster decode on Qwen3.6-35B and Gemma 4 26B/12B at 32K and 118K context with zero regressions — Gemma 4 12B at 118K goes from 5.13 to 13.81 tokens per second. It's still under review, and a reminder that the cheapest performance wins are often routing, not rewriting.

Disney is beta testing AI search and discovery on ESPN and Disney+. ESPN Search lets a subset of fans ask sports questions in natural language and get answers, stats, and recommendations drawn from ESPN's content and data ecosystem, while Disney+ experiments with intent-based discovery — describe what you want ("summer vibes") in natural language or voice, and get recommendations that fit the moment rather than viewing history. It's a small-scale test, but it's how the biggest entertainment company in the world is choosing to put LLMs

What to watch: The OpenJDK Governing Board's permanent generative-AI policy — and whether other open-source foundations follow Oracle's lead.

Should open-source projects ban AI-generated code — or is that fighting the future? Tell us in the comments.

Sources: llama.cpp PR #26689

Kling 4.0 goes live with 30-second native video at 10-bit HDR

Kuaishou's Kling AI has spent two years becoming the video model people actually finish projects with. Its new flagship, Kling 4.0, went fully live on October 15 — and this round is less about prettier frames than about the two things that make a generated shot usable in a real edit: length and control. Kling 4.0 generates up to 30 seconds in a single native pass, outputs 10-bit HDR at 4K, 1080p and 21:9, and takes up to 15 reference assets in one task. That budget splits across ten images, fi

Brain organoids classified Braille at 83% — in a trio, not alone

Two stories today: human brain tissue on a chip decoded a tactile signal, and the data-centre company behind some of the largest AI contracts in Europe turns out to be six months behind on its flagship site. Human forebrain organoids grown on a microelectrode array classified Braille letters with 61 percent accuracy on their own and 83 percent when the responses of three organoids were combined, in a peer-reviewed study from the University of Bristol. The team took event-based tactile readings

Suleyman says Microsoft's base model beat OpenAI's on a quarter compute

Microsoft finished pre-training its biggest in-house model yet — then used it to make a point about the company it bankrolls. Microsoft AI completed pre-training of MAI-Base-2, a mixture-of-experts foundation model with 80 billion active and 2.8 trillion total parameters, and says the recipe is 4.2–12.3x more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base. The run itself is the number worth keeping: 16,000 Nvidia GB200 GPUs for 38 days across August and September, at 21% average model

Amazon's new Bee devices listen all day — for $11.99 a month

Amazon turned its always-listening Bee line into a two-device family on Thursday, and NVIDIA opened a 2.8-trillion-token Indic corpus recipe. The connecting thread: both are bets that the next AI money is in memory — yours, or your model's. Amazon put two new always-listening devices on sale, and both now lead to a subscription. Bee Frontier, the next-generation wristband that also clips to clothing, ships now at $74.99 in Onyx, Lavender and Sage — up from the $50 predecessor — with seven-day

Read more

Kling 4.0 goes live with 30-second native video at 10-bit HDR

Brain organoids classified Braille at 83% — in a trio, not alone

Suleyman says Microsoft's base model beat OpenAI's on a quarter compute

Amazon's new Bee devices listen all day — for $11.99 a month