OpenAI's GPT-6.2 Astra and Sol give much shorter, more direct answers than their predecessors, but they are also less willing to admit uncertainty.
OpenAI's GPT-6.2 Astra and Sol give much shorter, more direct answers than their predecessors, but they are also less willing to admit uncertainty.
Nvidia researchers argue that language models often don't need their full context windows—which are slow, expensive, and prone to errors—and propose an alternative.
A Buddhist monastery in Vermont called Maple wants to establish itself as Silicon Valley's spiritual authority while creating its own AI god.
OpenAI has released new prompt templates for GPT-6.2 Astra and Sol meant to fix common model behaviors such as overly brief answers, unnecessary follow-up questions, and excessive testing when coding.
Anthropic mathematician Levent Alpöge says he used Claude to prove the Riemann hypothesis, math's most famous open problem for 167 years, with a proof formalized in Lean that spans over 250,000 pages.
Thinking Machines has released Inkling 2 Mini, a so-called looped transformer. A looped transformer runs certain layers multiple times for each token. Inkling 2 Mini runs 12 of its 24 layers three times in a row. That lets the model reason internally for longer without generating more text, finding the right answer with fewer tokens, according to Thinking Machines. The open mixture-of-experts model has 26.4 billion parameters, with 5.4 billion active per token because of the loops, the company says. It processes text and images, has a 1-million-token context window, and is licensed under Apache 2.0.
Inkling 2 Mini leads all open models under 100 billion parameters on agentic coding tests, scoring 48.2 percent on DeepSWE v1.1 and 46.5 percent on the Vals Index v2.2. Artificial Analysis gives it 20 points on its Intelligence Index, tying it with Qwen3.8 27B at roughly six times lower cost and four times the speed.

Inkling 2 Mini runs locally on Macs with at least 32 GB of memory and on Nvidia's DGX Spark. You can download the weights and a research version with variable loops from Hugging Face. It is also available to try in the Tinker Playground.
US tech stocks are tumbling. The tech-heavy Nasdaq 100 fell 5.6 percent on Friday and dropped 10.4 percent for the week, its worst showing since the Covid-19 crisis. That leaves the index 20 percent below its October peak and in bear market territory, according to the Financial Times. Doubts over multibillion-dollar AI buildouts, the Nvidia scandal and questionable China deals, and political headwinds after Democrats' midterm win are driving the sell-off like a perfect storm.
Investors are also debating whether high valuations for chip and memory stocks are still justified given their cyclical nature. Meanwhile, high oil costs from the Iran war are fueling persistent inflation and keeping interest rates high.
Nvidia took the hardest hit on Friday, plunging 14 percent after reports that the SEC is reviewing its accounting. The US semiconductor index also dropped 10.7 percent, with losses across memory and chip stocks including Sandisk, Micron, and Broadcom. Anthropic fell 9.1 percent a day after its market debut.
Bytedance and several US universities introduce Axion 1.0, an AI agent that rivals top commercial models from OpenAI and Anthropic on benchmarks after just three training phases.
OpenAI has sharply tightened safeguards against extracting internal reasoning from GPT-6.2 Astra and GPT-6.2 Sol. In internal automated tests, attacks pulling protected reasoning from prior requests dropped from a 54.0 percent success rate on GPT-6 Astra to 0.2 percent on GPT-6.2 Astra. Attacks on active requests fell from 22.9 to 0.9 percent, OpenAI says. The company is now bringing selected improvements to GPT-6 Astra, too.
The defenses work on two levels. For individual requests, the systems combine targeted training, input and output filters for messages and tool calls, and API blocks that stop older models from reading protected reasoning. Against coordinated campaigns, OpenAI bans thousands to tens of thousands of accounts weekly.
| Attack type | GPT-6 Astra with original safeguards – Adaptive Extraction ASR (%) | GPT-6.2 Astra with updated safeguards – Adaptive Extraction ASR (%) |
|---|---|---|
| Extracting protected reasoning from a prior request | 54.0 | 0.2 |
| Extracting protected reasoning from the current request | 22.9 | 0.9 |
Attack success rate (ASR) in adaptive black-box red teaming against GPT-6 Astra and GPT-6.2 Astra, aggregated across six attacker models. Lower is better.
Anthropic is fighting systematic distillation attacks as well, also banning tens of thousands of accounts to curb the theft of internal reasoning, reportedly succeeding in 90 to 99 percent of cases. Anthropic has also called for a coordinated response from industry and policymakers.
Signal is restricting screenshots to stop AI agents from reading private chats. The app blocks screenshots on iOS 27 and triggers protection on supported Android versions when an AI assistant tries to access the screen. Signal warned Reuters of an "existential threat" to privacy.
OS-level AI like Apple's Siri AI and Google's Gemini Intelligence can capture screen content to complete user tasks. That undermines end-to-end encryption by letting agents pull decrypted messages straight from the display and process them externally.
Meta's AI agent Muse recently captured Signal content on Android and macOS through broad setup permissions, including accessibility tools and screenshots on macOS. Meta says this happened only with explicit user permission.
Signal already blocked Microsoft's Recall on Windows but calls its defenses incomplete: macOS screenshots can't be reliably blocked, and AI access to file systems and notifications remains a risk. Signal says Apple, Google, and Microsoft should make it easy to block AI agents from apps entirely.