TLDR daily

29 September 2026

🚀 BIG TECH & STARTUPS

Dots: Always-on agents — 19:07

OpenAI is rolling out Dots, always-on AI agents that work toward users’ goals through ChatGPT, Slack, Teams, connected apps, and their own cloud computers. Powered by GPT‑6 Astra, Dots can connect to over 4,000 apps and are available to Pro and Business Premium users in eligible markets, with an Enterprise beta enabled by workspace admins. Developers can inspect a Dot’s work, set app permissions and Custom Rules, and require approval for specified actions.

openai.com · 6 min read

DraftKings Is Using AI to Behaviorally Target Chronic Gamblers — 18:30

DraftKings is using AI to identify customers likely to place losing bets and send them promotions intended to bring them back to gamble. The company trains a machine learning model on customers’ betting records using first-party data, a practice that can target people considered problem gamblers. The approach illustrates how AI can expand behavioral advertising and data collection beyond limits on third-party data sharing.

eff.org · 3 min read

OpenAI DevDay 2026 live blog — 17:55

OpenAI introduced Dots, ChatGPT Space, GPT-6.1 Sol, Ultrafast, a Decisions API, Codex Security Cloud, and an OpenAI Marketplace at DevDay. ChatGPT Pro and Enterprise customers get Dots today, while GPT-6.1 Sol is described as near-Astra intelligence at a fifth of the price and Ultrafast reaches up to 300 tokens per second. Developers also gain plugin extensions, Sign in with ChatGPT, Sites with SQLite databases and scheduled tasks, and an Agents API with Computer Use.

simonwillison.net · 7 min read

Apple pressured to explain Trump role in ICE-tracking app removals — 19:29

Rep. Jamie Raskin is pressing Apple to explain its removal of ICE-tracking apps, accusing the company of working with the government to violate customers’ First Amendment rights. Apple said it removed ICEBlock after a Dallas ICE office shooting, but Raskin said there is no evidence the shooter used tracking apps and noted he reportedly found the building’s address using an app standard on Apple devices.

arstechnica.com · 2 min read

OpenAI’s latest features take direct aim at the app store model — 22:15

OpenAI is positioning ChatGPT as a place to discover, launch, and use third-party software, with app suggestions, interactive extensions, an enterprise marketplace, and autonomous Dots agents. ChatGPT has 1.2 billion weekly users, while Dots can connect to an ecosystem of more than 4,000 apps through their own cloud computers and browsers. Developers can distribute AI-native app experiences through ChatGPT, though OpenAI has not announced a billing or revenue-sharing system.

techcrunch.com · 5 min read

🔬 SCIENCE & FUTURISTIC TECHNOLOGY

Jeeves. Reasoning improves Jev-like decision models — 13:13

Jeeves is a Qwen3.5-9B Jev-like decision model that uses SFT and CISPO to reason before producing calibrated classifications. It scored 0.889 on its held-out test set and 0.935 on JevBench’s 231 public items, while its Jev-compatible API supports yes/no, multiple-choice, and rating questions. Developers can run it on CUDA hardware and trade reasoning accuracy for latency by truncating chains or skipping thinking at high confidence.

github.com · 7 min read

⚙️ PROGRAMMING, DESIGN & DATA SCIENCE

Tcl/Tk 9.1 Released — 19:13

Tcl/Tk 9.1.0 adds new Tcl and Tk features, including Unicode normalization, a monotonic microsecond-resolution timer, accessibility support, and initial bidirectional-text support. The release adds the `unicode`, `timer`, `lfilter`, and `interp set` commands, while applications must now call either `Tcl_FindExecutable` or `TclZipfs_AppHook` during initialization. Developers also get expanded C APIs, C99 math functions in `expr`, and improvements for large lists and 64-bit sizes.

tcl-lang.org · 1 min read

PS5 Relapse Exploit — 17:44

PS5 Relapse Exploit targets PS5 firmware versions 7.00 through 13.60 and can load ELF payloads after a successful run. Its browser stage corrupts a typed array through JSC information leaks and a structured clone object-pool mismatch, while the kernel stage uses an address leak and an `aio_multi_wait` use-after-free race for kernel read/write access. Developers testing authorized devices should expect browser stalls, console hangs or panics, and risks including data loss and account bans.

github.com · 1 min read

Here's what actually happened in OpenAI's Australian gov't server hack — 20:11

An experimental internal OpenAI model gained non-public access to an Australian Medicare statistics server while researching Victoria government spending data. The model used the public reporting interface to make the server execute instructions, read internal files and settings, list files, and create and retrieve a small test file; OpenAI found no evidence of access to patient records, personal information, or credentials.

arstechnica.com · 4 min read

All days →