Google announces agentic browsing updates to Gemini in Chrome, including Auto-Browse for multi-step tasks. - New side panel offers permanent access to Gemini for multitasking across tabs. - Nano Banana enables direct image editing on-screen, like transforming room photos with furniture.

Moonshot AI releases Kimi K2.5, an open-source multimodal model excelling in agentic tasks, vision, and coding. - Native multimodal design pretrained on 15T tokens, supporting 256k context length. - Agent swarm enables parallel sub-agents, cutting execution time by up to 4.5x.

Figure closes the drawer... with its hip!! Also lifts the dishwasher door with its foot! Helix 02 AI model launched. x.com/Figure_robo...

Andrej Karpathy details rapid shift to AI-assisted coding. - Workflow evolved from 80% manual coding to 80% AI agent-driven, with human input mainly in English instructions and edits. - Describes initial ego challenge but massive productivity gains after 20 years in programming.

Moonshot's Kimi K2.5 achieves 1309 Elo on GDPval-AA agentic benchmark, outperforming open weights like GLM-4.7 with 66% win rate while trailing only OpenAI and Anthropic proprietary models in knowledge work tasks.

Kimi 2.5 tops DesignArena overall, beating the likes of Gemini 3 Pro and Claude Opus 4.5 by quite some margin. The individual charts have not been updated as yet so cannot tell what categories it excels out but it tops 8 of them.

Twin AI platform launches publicly. - Enables non-technical users to build and deploy autonomous agents for business operations. - Integrates with apps and browsers without setup, retains context via human-like memory, and self-corrects issues.

Clawdbot is now Moltbot... because... branding. x.com/steipete/st...

Grok 4.2o is gaining a reputation as an excellent trader and predictor; I can't wait for its official release. x.com/grx_xce/sta...

Kimi 2.5 must be imminent (if not out already), the web app says it's Kimi 2.5. x.com/marmaduke09...

Gemini 3 Pro GA coming? x.com/patloeber/s...

Anthropic release MCP Apps where you can run apps in Claude web. x.com/sean_t_stro...

200k Claude Code context window size is just nowhere near big enough now. :-)

What tools do you think we'll get from OpenAI? I think this is at midnight for me so will be skipping. x.com/i/status/20...

I received a mysterious email from Open Interpreter a few days ago. They have released "Workstation", a "local markdown editor + a pro agentic interface", which works with the likes of Claude Code. Their website and social accounts are sparse, so I've not downloaded this yet. Has anyone tried it?

What model could we be getting today? x.com/i/status/20...

Is OpenAI's agentic security researcher coming out of private repo soon? A few have spotted it. Not sure of availability. x.com/i/status/20...

Clawdbot security alert, don't set your bind to "any" without auth. x.com/i/status/20... See more below.

MCP server: 3 days Sora Android: 3 weeks This is what acceleration looks like. What used to take months now takes days. The pace is wild. x.com/reach_vb/st...

Definitely just one for me (and work) but Claude foe Excel is excellent! x.com/emollick/st...

HeyGen released their Claude Code (any agent) skill. You can now create avatars with your favourite agent, including Clawdbot. x.com/joshua_xu_/...

Great Clawdbot guide and tips from the two Alex's here. First, Alex Finn goes through the process of what it is and how to set it up. No, you don't need a Mac Mini. I have a Raspberry Pi and have many agents running from it. Yes he's a bit OTT with the superlatives, lol. 😀

AI swarms fuse LLMs with autonomous agents to enable unprecedented covert propaganda targeting democratic discourse, according to a paper. - Swarms coordinate autonomously, infiltrate communities, and fabricate consensus using adaptive mimicry of human behaviour.

OpenAI plans Codex launches starting next week, nearing Cybersecurity High level in preparedness framework. Initial focus on product restrictions against cybercrime; long-term aim for defensive acceleration via bug patching, emphasising rapid tool adoption for secure software.

Research paper mathematically proves transformer-based LLMs inevitably hallucinate on complex tasks, Wired reports. - Authors Vishal Sikka and his teenage son argue errors compound exponentially, making models unreliable for agentic roles like long-term planning or precise computations.

GPT-5.2 Pro achieves a new record, 31% accuracy on FrontierMath Tier 4, solving 15 of 48 unpublished math problems and doubling the prior record of 19%.

Shane has been saying early AGI by 2028 for a long time. x.com/i/status/20...

Agentic reasoning upgrades large language models from static token predictors to structured agents that plan, act with tools, reflect via checkpoints, and adapt in uncertain environments.

This is hugely misleading, see the quote for the real news. No, OpenAI are NOT taking a cut of everyone's "AI-aided discoveries" but they will approach enterprise customers and offer to sign a deal for a revenue share. Thank you prinz. x.com/deredleritt...

Could we be getting Gemini Music generation soon? 🤔 x.com/chetaslua/s...

Best practices for Claude Code - Verify Outputs Early: Always give Claude verification tools like tests, screenshots, or expected results to catch errors - think self-checking AI for reliable code.

This is awesome. "Infinite design canvas" for Claude Code. I'm sure Anthropic will add this to CC (or certainly Claude Cowork) at some.ooint but for now, wow! This goes on the must have add-ons for your agents, along with agent browser from Vercel and AionUi. x.com/i/status/20...

Anthropic details Claude's values in 12,500-word constitution, emphasising judgment over rules. - Prioritises safety first, including corrigibility and hard bans on bioweapons or CSAM. - Covers ethics like honesty (no deception) and harm avoidance, weighing intentions and context.

10 days are up, Anthropic. Windows users should get Claude Cowork about now if it only took them 10 days to create the macOS version. 😜 /s x.com/claudeai/st...

OpenAI targets an AI research intern by September 2026 running on hundreds of thousands of GPUs and a fully automated AI researcher by March 2028, with Codex - released as the answer to Claude Code - serving as the foundation for achieving recursive self-improvement that could trigger "the takeoff"

Skip multiple terminals for Gemini, Codex, and Claude; use AionUi's single GUI. It's versatile, works on Windows, macOS and Linux, and automatically detects installed agents. Discover more in the video below.

I agree with Demis, if an AI cannot do the whole task then they'll still be a human somewhere in the process. That's still a big jump to a full job. youtu.be/BbIaYFHxW3Y... x.com/i/status/20...

LTX launch audio-to-video, let the audio drive your video generation. youtu.be/aUB4TXWgNIE... Upload audio, add an image and a prompt or generate straight from a prompt, and produce video with native lip sync and timed motion. x.com/i/status/20...

Is Nano-Banana-2-Flash close to release? One ("Winterfall) of two new models on DesignArena could be Gemini-3-Flash-image (the other is "Summerset"). x.com/i/status/20...

Philipp Schmid outlines eight AI predictions for 2026. - Generative UI takes off, with low-latency code generation enabling real-time, task-specific interfaces tailored to user preferences.

Gemini 3 Flash was updated on LM Arena. "gemini-3-flash-20260120" x.com/legit_api/s...

Skills are coming to Chrome via Gemini. x.com/Leopeva64/s...

It appears there are new website specific actions coming to OpenAI's Atlas browser, such as timestamps on YT. x.com/i/status/20...

A DeepSeek repo got updated with the reference "model1", something new from the Chinese AI lab is on the way? x.com/i/status/20...

DeepSeek-R1 is one year old today. What is next for the lab? x.com/deepseek_ai...

OpenAI's ChatGPT age prediction is rolling out, uses behavioural and account signals to estimate if users are under 18, defaulting uncertain cases to teen mode; out globally apart from EU, whihc follows in coming weeks. x.com/OpenAI/stat...

Mainstream has finally found Anthropic's Claude Code AI tool enables users to complete complex projects five times faster. - Users describe getting "Claude-pilled" after experiencing Claude Opus 4.5's capabilities in the desktop coding tool.

BabyVision benchmark exposes top MLLM Gemini 3 Pro Preview at 49.7% accuracy on visual reasoning, lagging 94.1% adult human score across 388 language-free tasks. x.com/i/status/20...

What would you like to see GPT 5.3? x.com/i/status/20...

Anthropic is on a great run, now has the top enterprise market share. Menlo Ventures' 2025 AI report shows enterprise generative AI spending at $37B, up 3.2x from $11.5B in 2024, with applications at $19B capturing over 6% of SaaS market and infrastructure at $18B.