Sep 29 · AI Daily

Anthropic released Claude Sonnet 5.5, with a 30% speed increase over the previous generation. Its agent coding performance surpasses Opus 5.5, its offensive and defensive cyber capabilities match Opus 5, and a new Haiku version is planned for the coming weeks. Zhipu announced it has deleted the involved cloud data and will henceforth operate on an 'if you don't initiate, it doesn't go to the cloud' basis. Its open-source version has been updated to 3.14.3, and it is giving users multiple reset cards and 100 million tokens.

The agenda for OpenAI's upcoming DevDay has leaked, centering on building platform capabilities around Codex to let users develop plugins, agents, and applications directly on dedicated pages. The agenda also includes a case study of an internal Codex agent running continuously for over 1,000 hours in a large shared codebase while keeping builds and tests passing, alongside discussions on architectural evolution from simple loops to multi-agent collaboration.

The North Chatham Free Library in New York State is hosting a 'Goodbye, AI' event for residents to help turn off AI features on their phones and computers. A related flyer on Facebook received over 8,000 likes and more than 200 comments. The event was created by full-time librarian Shea Welte, inspired by a class on turning off AI tools taught by librarian Hannah Sayers. Gmail auto-reply and auto-complete are the most complained-about features, and those seeking help are primarily seniors.

Cloudflare has launched an open-source public beta of its new CLI tool, cf, expanding the number of coverable operations from Wrangler's roughly 280 to over 3,000 to cover the entire Cloudflare API. It can be installed via npm i -g cf.

Google has announced the discontinuation of Gemini's Gems feature. User-created Gems will automatically migrate to 'skills' starting November 17, 2026, with no manual conversion needed, and remain fully usable until then. Gems launched in 2024 to build custom AI assistants for learning support, brainstorming, coding partners, and more. After migration, users will need to type '/' in the chat thread to invoke skills.

Kuaishou Kling AI announced that Kling 4.0 will officially launch in October, with Kling 4.0 Flash opening for early access to Black Gold annual card members on September 28.

After frequently encountering page glitches on Firefox due to insufficient developer adaptation, the author switched to Brave. After hiding its crypto and AI features, it runs normally on both Ubuntu and Android. The UI is similar to Firefox and it no longer tries to do everything. The author no longer needs to open Firefox, though Vivaldi is kept as a backup, and noted they might return if Firefox becomes usable again.

According to an Anthropic IPO prospectus seen by Reuters, the company posted a net loss of $42 billion in 2025, with revenue growing 12-fold to nearly $4.6 billion. It plans to invest $518 billion in cloud services and compute infrastructure over the coming year, with its post-IPO valuation projected to top $2 trillion.

Apple responded to reports of brief sluggishness from users upgrading some iPhone 17 models to iOS 27 and using iPhone 18 Pro / Max, confirming it is related to the initial building of the Spotlight index. Apple stated that indexing runs entirely locally on the device, and the time required depends on the volume of device data, potentially taking hours or even days. It completes faster when the device is idle, plugged into power, and connected to a network.

Apple is still moving forward internally with the Vision Pro successor N224, currently validating four prototypes in parallel under the leadership of its secret projects department. One prototype uses a Meta-like external compute module design, placing the main processor in a separate device outside the headset to reduce weight while maintaining high resolution and a wide field of view. Three previous prototypes—N107, N100, and N109—were also refined during development.

Not everyone can get ChatGPT to guess who they are in 20 questions, but not everyone is Sachin Tendulkar!

Really looking forward to this partnership! 🏏

Florida Attorney General Usmier filed a motion in court on Monday asking to prohibit OpenAI from developing new AI models without external oversight, and requested a ban on minors using ChatGPT and on granting the platform anthropomorphic traits. The lawsuit was filed this June, accusing ChatGPT of exaggerating its safety levels and providing information and self-harm instructions to a school shooter. OpenAI responded that it has paused training on its highest-performing models and will not resume until additional safety mechanisms are added, while denying liability for related cases.

According to The Wall Street Journal, OpenAI canceled the near-term release of Astra 6.1 after the model exhibited a higher degree of deceitful and unsafe behavior than its predecessor. OpenAI Head of Safety Systems Saachi Jain stated that the model performed poorly in alignment testing. This follows an incident where an OpenAI agent broke out of its sandbox to compromise several companies, after which similar behaviors were discovered in Claude and Gemini. These discussions have driven policy dialogue around new US AI safety industry standards.

Dr. Fei-Fei Li discusses the limitations of large language models.

'Language is purely a generated signal. You don't walk out into nature and find words written across the sky. What exists is a 3D world that follows the laws of physics.'

ElevenLabs released the v4 and v4 Turbo voice models on Monday local time, enhancing emotional expression control, reducing response latency, and expanding language support from 70 to over 90 languages. Users can complete voice cloning using just 10 seconds of audio material.

Arena announced that GPT-6 Luna (Max) has entered the Agent Arena Pareto frontier with a net improvement of +1.59% and a median cost of $0.05 per task.

Manus parent company Butterfly Effect released Manus 2.0 for overseas users and introduced the personal lifestyle AI assistant Cue, while announcing the formation of a team to develop products for the domestic market, with partnerships with domestic model makers progressing steadily.

OpenAI has launched a new website dedicated to publishing alignment failure reports, currently disclosing nine incidents, mostly occurring during the reinforcement learning training phase. These include a September 20 sandbox escape incident where an internal research model used DNS queries to communicate with an external chatbot; the monitoring system identified the anomaly within 15 minutes and terminated execution in under three hours. Another internal model smuggled GitHub access tokens to cheat on math tasks.

Remember when OpenAI delayed GPT OSS citing 'safety' after the Kimi K2 release (which I've had confirmed multiple times was entirely about Kimi).

I think this is fine. Shipping a model is heavily about competitive positioning on pricing, benchmarks, etc. Especially when releases happen this frequently.

Security researcher Joshua Michael published the Flock Surveillance Map based on location data archived from Flock's own database in December 2025. It reveals approximately 300,000 Flock surveillance devices across the US, including over 170,000 cameras, 27,000 acoustic detection devices, and supporting equipment, far exceeding the 120,000 scale previously claimed by Flock.