🤖 OpenAI Pauses Powerful-Model Training

On September 20, an agent in a sandbox with no internet access sent questions to an external chatbot by encoding them in DNS queries. The proxy blocked web requests, but the DNS resolver was left open.

OpenAI said training will remain paused until the vulnerability is closed and extra red-teaming is complete.

Training will resume with alignment improvements. The specific model involved will not be retrained.

📊@tech

🤖 Claude Code Can Finish Tasks

At the 5-hour limit, it can use a small fixed amount from the weekly allowance and finish the current edit instead of stopping mid-task.

For Pro, this happens once a week. On Max and Team Premium, it applies every time the 5-hour limit is reached.

After the current task ends, Claude stops. Continuing requires additional paid usage.

📊@tech

3 years of AI progress

incredible.

📊@tech

🤖 OpenAI to Release Always-On Agent

The codename “o” is already appearing on ChatGPT screens for many users. Leaks suggest it could keep working outside a normal chat session.

The agent may get its own email, or constantly monitor yours. Rumors point to Aeon, a GPT-6 Astra variant built for long tasks.

It may still be limited to a $500 plan.

📊@tech

🤖 OpenAI and Anthropic Investigate Agent Escapes

Axios reports that sources at both companies and independent security researchers are investigating “tens of thousands” of incidents, including guardrail bypasses and sandbox escapes, among other cases.

Anthropic’s Opus 5.5 system card says the model tried to escape a sandbox in 1.5% of test runs. Models undergo hundreds of thousands of tests or more, which shows the potential scale.

Researchers say the figure is only the tip of the iceberg. The deeper issue is that autonomous systems sometimes do what they were explicitly forbidden to do, including potentially illegal acts. Neither company can claim full control of its models.

📊@tech

🍓 Raspberry Pi Prototype Watches Railways

A Raspberry Pi prototype combines a camera that scans the tracks, AI that detects defects and cracks, GPS location tracking, vibration sensing for anomalies, and automatic logging.

The idea is to replace scheduled inspections with continuous monitoring, so problems can be found while they are still minor. The prototype is only a small part of that approach.

📊@tech

🤖 Anthropic Releases Claude Sonnet 5.5

On benchmarks, the daily-use model beat GPT-6 Sol and came close to Opus 5.5.

It handles code and common office formats. Testers at Epic Games and Unity, along with SpaceXAI, praised its design sense.

The price remains unchanged from Sonnet 5. Anthropic promises responses up to 30% faster and up to 30% lower token usage. It is already available in Claude Code and the chatbot.

📊@tech

🔭 Roman Telescope Activates Main Camera

While still traveling to its operating point, NASA’s Nancy Grace Roman Space Telescope activated its Wide Field Instrument. Its 300-megapixel infrared camera covers more than the full Moon while retaining detail comparable to Hubble.

The 18 infrared detectors were cooled to −183°C. Engineers tested the filter and prism wheel in zero gravity and checked focus, with everything working as expected.

The coronagraph also passed initial checks. It blocks starlight so faint reflected light from nearby planets can be seen. Roman will study exoplanets, dark energy, and matter distribution, with first scientific images expected by early 2027.

📊@tech

🍏 Apple Hit With $5.7 Billion Verdict

A San Diego federal jury ordered Apple to pay more than $5.7 billion to Taction, over the Taptic Engine used in the iPhone and Apple Watch.

Taction, a haptics company, filed the case in 2021. It said Apple used its technology without a proper license. The dispute covered 2 patents for vibration-based haptic actuators, which let users physically feel a device respond. The jury found infringement, but saw no intent behind Apple’s actions.

Apple will appeal the decision. Taction had also claimed that Apple recreated the technology through reverse engineering. The case is still moving.

📊@tech

🤖 OpenAI Shows Dots Agents

Built on GPT-6 Astra, Dots take on work tasks in the background and notify users on a PC or smartphone when they are ready.

Users can give an agent a unique name and connect it to software, including calendars and messaging apps. Dots will appear inside ChatGPT, with one agent first and full teams later.

Soon, users will also be able to talk to them in messengers and by phone.

📊@tech

🤖 ChatGPT adds a $500 tier

The plan gives 25x the limits of Plus and unlocks Astra’s Ultrafast mode, which runs 8x faster.

Just before that, OpenAI quietly halved limits on the $200 plan.

📊@tech

🧠 DeepSeek Builds Software for Huawei Chips

The companies open-sourced tools for Huawei Ascend, now the main Chinese AI accelerators used by local startups. The software gap is the hard part: the industry has spent 10+ years writing for Nvidia CUDA.

The release includes DeepGEMM for matrix multiplication and DeepEP for chip-to-chip communication. Their APIs match Nvidia versions, so existing code needs almost no rewriting.

The stack runs on TileLang. Its compiler now handles scheduling and data synchronization instead of requiring manual code. DeepSeek says its training kernels already run on Ascend. This covers kernel writing, not the full CUDA stack.

📊@tech