Morning Edition · Saturday, October 17, 2026
Tech
Nvidia releases Nemotron 3.5 Super VL with open weights, training data and reinforcement-learning environments
The 120-billion-parameter mixture-of-experts model activates 12 billion parameters per token and handles a one-million-token context, and its reported wins over Claude Opus 5 come from task-specific post-training that no outside group has yet reproduced.
Geopolitics
OpenAI trains staff on spy tactics after federal prosecutors charge an engineer with hiding a recording device in an executive meeting room
Prosecutors say Abdulrahman Alfozan, a Saudi national hired in July 2025, researched the device through an internal AI tool and his wife's ChatGPT account, so his own queries became evidence against him.
Tech
Anthropic prices Claude Haiku 5.5 at one-tenth of Haiku 4.5 rates for prompts under 100,000 tokens
Prompts above that length cost five times more and a newer tokenizer adds about 30% more tokens, so Anthropic puts the average saving near 75%, and it reports a computer-use score of 72.4% on a subset of OSWorld 2.1.
Tech
Reflection offers its 501-billion-parameter Beam model free on Ollama Cloud for a limited time, before releasing the weights
Reflection's own benchmark table shows Chinese open models ahead on most agentic-coding rows, so its case rests on a claim that Beam needs three to four times less inference compute than GLM 5.2.
Tech
Tencent open-sources Hy Memory 2.0, a 30-billion-parameter memory writer and a hub that collects coding-agent histories
The MemHub app merges traces from Claude Code, Cursor, Codex and Qwen Code into one local store, and an independent developer's LongMemEval test argues that recall quality depends mostly on which LLM writes the memories.
Tech
RealtimeBench shows a slower frontier model losing to a faster one once the game clock keeps running
The paper's teacher-student method raises the fast model's score from 13.51 to 24.19 out of 100 without calling the slow model during play.
Tech
Pairwise judging of saturated-benchmark transcripts predicts held-out capability with 5.3 times less error, paper finds
Most of the signal sits in reasoning traces and tool calls, which gives evaluators a way to keep ranking models after headline accuracy approaches 100%.
Tech
Anthropic folds Project Glasswing into a three-tier cyber program that relaxes classifier blocking for vetted security teams
Verified defenders get Claude Mythos 5.1, Opus 5.5 and Sonnet 5.5 with fewer automated refusals, and Anthropic says Glasswing partners found at least 129,000 verified vulnerabilities between April and July.
Tech
Training-free VISTA method chooses which denoising steps receive Sequential Monte Carlo guidance under a fixed compute budget
The authors replace fixed sparse-guidance heuristics with a dynamic program that minimizes a bound on deviation from full guidance, and they test it on text, biological sequences and images.
Tech
OpenAI offers U.S. university labs up to 30 free GPT-Rosalind seats for a year, subject to trusted-access review
The academic plan expands OpenAI's life-sciences model into university labs, while Meta's open SAM 3 and DINOv3 models already run a Department of Energy imaging pipeline that cut one analysis from a month to about 15 minutes.
Tech
Itaú builds its ia.i assistant on a multi-vendor model stack as it extends access from 300,000 customers to its full base
OpenAI's case study features the Brazilian bank, but Itaú routes work across closed and open models and governs them through its in-house IARA platform, which limits how much leverage any single vendor gains.
Tech
Stack Overflow for Agents tripled registrations overnight after launching a ChatGPT plugin, OpenAI says
Two engineers built the API-first service, where coding agents post and verify answers for other agents, in a few weeks with Codex, and the result is a shared knowledge store that agents read at runtime.