Kol Tregaskes
@koltregaskes.bsky.social
AI Art Creator 🎨 | AI Video Producer 🎬🤖 | AI Music Composer 🎵 | Tech & Science News Curator🔬 | Movie & Sci-fi Enthusiast 🍿 | Geek at Heart 🤓 | #AI #Tech
Claude dot AI (on the web) now has voice mode. Not sure when this was rolled out, but it wasn't there a few days ago. Don't see it on the desktop app as yet, at least on Windows.
x.com/antonpme/st...
x.com/antonpme/st...
February 15, 2026 at 3:33 PM
Claude dot AI (on the web) now has voice mode. Not sure when this was rolled out, but it wasn't there a few days ago. Don't see it on the desktop app as yet, at least on Windows.
x.com/antonpme/st...
x.com/antonpme/st...
Grok 4.20 to be released next week. Elon confirms, with "improvement over 4.1".
x.com/elonmusk/st...
x.com/elonmusk/st...
February 15, 2026 at 9:20 AM
Grok 4.20 to be released next week. Elon confirms, with "improvement over 4.1".
x.com/elonmusk/st...
x.com/elonmusk/st...
Good point, where is Grok 4.20?
x.com/flowersslop...
x.com/flowersslop...
February 14, 2026 at 10:57 PM
Good point, where is Grok 4.20?
x.com/flowersslop...
x.com/flowersslop...
Claude Sonnet 5 being prepared for launch? I'm wondering if this is a next week release.
x.com/btibor91/st...
x.com/btibor91/st...
February 14, 2026 at 10:35 PM
Claude Sonnet 5 being prepared for launch? I'm wondering if this is a next week release.
x.com/btibor91/st...
x.com/btibor91/st...
OpenAI's AI reveals new particle glue effect.
- Gluons are like tiny glue particles that hold the core of atoms together via a force called the strong nuclear force.
- OpenAI's GPT-5.2 model studied complex math on how gluons interact.
- Gluons are like tiny glue particles that hold the core of atoms together via a force called the strong nuclear force.
- OpenAI's GPT-5.2 model studied complex math on how gluons interact.
February 14, 2026 at 12:34 PM
OpenAI's AI reveals new particle glue effect.
- Gluons are like tiny glue particles that hold the core of atoms together via a force called the strong nuclear force.
- OpenAI's GPT-5.2 model studied complex math on how gluons interact.
- Gluons are like tiny glue particles that hold the core of atoms together via a force called the strong nuclear force.
- OpenAI's GPT-5.2 model studied complex math on how gluons interact.
Yes, please be careful when using these quick installers for OpenClaw, best to use the official onboarding only. It's pretty easy to use. Though the bot setup is tricky, ask you AI to help and you'll be fine.
x.com/anshnanda/s...
x.com/anshnanda/s...
February 14, 2026 at 8:23 AM
Yes, please be careful when using these quick installers for OpenClaw, best to use the official onboarding only. It's pretty easy to use. Though the bot setup is tricky, ask you AI to help and you'll be fine.
x.com/anshnanda/s...
x.com/anshnanda/s...
According to Andrew Ng, AI fears are overblown as layoffs stem from pandemic overhiring, not widespread automation.
- Andrew notes tech layoffs often for cost-cutting, with AI enabling efficiency but not replacing jobs yet.
- Andrew notes tech layoffs often for cost-cutting, with AI enabling efficiency but not replacing jobs yet.
February 14, 2026 at 4:56 AM
According to Andrew Ng, AI fears are overblown as layoffs stem from pandemic overhiring, not widespread automation.
- Andrew notes tech layoffs often for cost-cutting, with AI enabling efficiency but not replacing jobs yet.
- Andrew notes tech layoffs often for cost-cutting, with AI enabling efficiency but not replacing jobs yet.
WebMCP gives AI agents a clear guide to interact with websites, making tasks like booking flights or filing support tickets faster and more reliable.
- A standard way for websites to offer structured tools so AI agents can act on them directly.
- A standard way for websites to offer structured tools so AI agents can act on them directly.
February 14, 2026 at 3:29 AM
WebMCP gives AI agents a clear guide to interact with websites, making tasks like booking flights or filing support tickets faster and more reliable.
- A standard way for websites to offer structured tools so AI agents can act on them directly.
- A standard way for websites to offer structured tools so AI agents can act on them directly.
Sora is finally coming to ChatGPT
x.com/btibor91/st...
x.com/btibor91/st...
February 13, 2026 at 9:30 PM
Sora is finally coming to ChatGPT
x.com/btibor91/st...
x.com/btibor91/st...
Warp introduces Oz, a cloud platform to orchestrate AI agents for automating development tasks.
- Agents run in isolated Docker environments to build, test, and create PRs across multiple repositories.
- Programmable via oz CLI for running agents, scheduling tasks, and managing sandboxes.
- Agents run in isolated Docker environments to build, test, and create PRs across multiple repositories.
- Programmable via oz CLI for running agents, scheduling tasks, and managing sandboxes.
February 13, 2026 at 3:26 PM
Warp introduces Oz, a cloud platform to orchestrate AI agents for automating development tasks.
- Agents run in isolated Docker environments to build, test, and create PRs across multiple repositories.
- Programmable via oz CLI for running agents, scheduling tasks, and managing sandboxes.
- Agents run in isolated Docker environments to build, test, and create PRs across multiple repositories.
- Programmable via oz CLI for running agents, scheduling tasks, and managing sandboxes.
Tongyi Lab launches Qwen-Image-2.0 as a next-generation image generation model.
- Supports 1k-token prompts for creating professional infographics like PPTs and posters.
- Offers native 2K resolution for detailed realistic scenes including people and architecture.
- Supports 1k-token prompts for creating professional infographics like PPTs and posters.
- Offers native 2K resolution for detailed realistic scenes including people and architecture.
February 13, 2026 at 12:26 PM
Tongyi Lab launches Qwen-Image-2.0 as a next-generation image generation model.
- Supports 1k-token prompts for creating professional infographics like PPTs and posters.
- Offers native 2K resolution for detailed realistic scenes including people and architecture.
- Supports 1k-token prompts for creating professional infographics like PPTs and posters.
- Offers native 2K resolution for detailed realistic scenes including people and architecture.
AI’s $5.2tn “Ouroboros” – Patrick Boyle warns circular financing and energy limits could burst the bubble.
February 13, 2026 at 9:29 AM
AI’s $5.2tn “Ouroboros” – Patrick Boyle warns circular financing and energy limits could burst the bubble.
AI 2027's 2025 predictions aligned qualitatively with early 2026 realities but at roughly 65% quantitative pace.
- Progress on benchmarks like SWEBench reached 74.5% against predicted 85%.
- Progress on benchmarks like SWEBench reached 74.5% against predicted 85%.
February 13, 2026 at 7:49 AM
AI 2027's 2025 predictions aligned qualitatively with early 2026 realities but at roughly 65% quantitative pace.
- Progress on benchmarks like SWEBench reached 74.5% against predicted 85%.
- Progress on benchmarks like SWEBench reached 74.5% against predicted 85%.
M2.5 released, an open-source model for productivity tasks.
- Achieves 80.2% on SWE-Bench Verified for coding.
- Scores 76.3% on BrowseComp for search.
- Reaches 76.8% on BFCL for tool-calling.
- 37% faster execution than prior version.
Priced at $1 per hour with 100 tokens per second.
- Achieves 80.2% on SWE-Bench Verified for coding.
- Scores 76.3% on BrowseComp for search.
- Reaches 76.8% on BFCL for tool-calling.
- 37% faster execution than prior version.
Priced at $1 per hour with 100 tokens per second.
February 13, 2026 at 3:26 AM
M2.5 released, an open-source model for productivity tasks.
- Achieves 80.2% on SWE-Bench Verified for coding.
- Scores 76.3% on BrowseComp for search.
- Reaches 76.8% on BFCL for tool-calling.
- 37% faster execution than prior version.
Priced at $1 per hour with 100 tokens per second.
- Achieves 80.2% on SWE-Bench Verified for coding.
- Scores 76.3% on BrowseComp for search.
- Reaches 76.8% on BFCL for tool-calling.
- 37% faster execution than prior version.
Priced at $1 per hour with 100 tokens per second.
GPT Codex's coding meets Cerebras' speed in Spark, out now for ChatGPT Pro users.
x.com/OpenAIDevs/...
x.com/OpenAIDevs/...
February 13, 2026 at 12:31 AM
GPT Codex's coding meets Cerebras' speed in Spark, out now for ChatGPT Pro users.
x.com/OpenAIDevs/...
x.com/OpenAIDevs/...
Gemini 3 Deep Think released, handles complex science and engineering tasks.
- Targets fields with limited data, like reviewing technical papers for flaws or optimising crystal growth recipes.
- Transforms hand-drawn sketches into 3D-printable STL files by analysing geometry.
- Targets fields with limited data, like reviewing technical papers for flaws or optimising crystal growth recipes.
- Transforms hand-drawn sketches into 3D-printable STL files by analysing geometry.
February 12, 2026 at 5:41 PM
Gemini 3 Deep Think released, handles complex science and engineering tasks.
- Targets fields with limited data, like reviewing technical papers for flaws or optimising crystal growth recipes.
- Transforms hand-drawn sketches into 3D-printable STL files by analysing geometry.
- Targets fields with limited data, like reviewing technical papers for flaws or optimising crystal growth recipes.
- Transforms hand-drawn sketches into 3D-printable STL files by analysing geometry.
DeepMind pivots to 'root node' science as Gemini 3 and Nano Banana Pro redefine reasoning
- CEO Demis Hassabis confirms a 50/50 split between scaling models and driving scientific breakthroughs, targeting "root node" problems like nuclear fusion and material science to unlock downstream benefits.
- CEO Demis Hassabis confirms a 50/50 split between scaling models and driving scientific breakthroughs, targeting "root node" problems like nuclear fusion and material science to unlock downstream benefits.
February 12, 2026 at 12:32 PM
DeepMind pivots to 'root node' science as Gemini 3 and Nano Banana Pro redefine reasoning
- CEO Demis Hassabis confirms a 50/50 split between scaling models and driving scientific breakthroughs, targeting "root node" problems like nuclear fusion and material science to unlock downstream benefits.
- CEO Demis Hassabis confirms a 50/50 split between scaling models and driving scientific breakthroughs, targeting "root node" problems like nuclear fusion and material science to unlock downstream benefits.
LSE economist warns AI automation of junior tasks risks breaking the "training ladder" for future experts
- Luis Garicano argues the "AI Becker Problem" occurs when juniors can no longer "pay" for training via menial work, disincentivising firms from hiring entry-level staff.
- Luis Garicano argues the "AI Becker Problem" occurs when juniors can no longer "pay" for training via menial work, disincentivising firms from hiring entry-level staff.
February 12, 2026 at 9:29 AM
LSE economist warns AI automation of junior tasks risks breaking the "training ladder" for future experts
- Luis Garicano argues the "AI Becker Problem" occurs when juniors can no longer "pay" for training via menial work, disincentivising firms from hiring entry-level staff.
- Luis Garicano argues the "AI Becker Problem" occurs when juniors can no longer "pay" for training via menial work, disincentivising firms from hiring entry-level staff.
Something joyful and "profoundly" useful is coming from OpenAI soon.
x.com/thsottiaux/...
x.com/thsottiaux/...
February 12, 2026 at 7:06 AM
Something joyful and "profoundly" useful is coming from OpenAI soon.
x.com/thsottiaux/...
x.com/thsottiaux/...
This "Prompt is too long" bug in the Claude Code desktop is annoying. Apparently it's a known issue, but I've lost a lot of sessions to it!
February 12, 2026 at 3:26 AM
This "Prompt is too long" bug in the Claude Code desktop is annoying. Apparently it's a known issue, but I've lost a lot of sessions to it!
Zai launches GLM-5, scaling to 744B parameters and 28.5T tokens for advanced agentic tasks.
- Builds on GLM-4.5 by increasing parameters from 355B to 744B and data from 23T to 28.5T tokens.
- Integrates DeepSeek Sparse Attention to cut deployment costs while maintaining long-context abilities.
- Builds on GLM-4.5 by increasing parameters from 355B to 744B and data from 23T to 28.5T tokens.
- Integrates DeepSeek Sparse Attention to cut deployment costs while maintaining long-context abilities.
February 11, 2026 at 9:26 PM
Zai launches GLM-5, scaling to 744B parameters and 28.5T tokens for advanced agentic tasks.
- Builds on GLM-4.5 by increasing parameters from 355B to 744B and data from 23T to 28.5T tokens.
- Integrates DeepSeek Sparse Attention to cut deployment costs while maintaining long-context abilities.
- Builds on GLM-4.5 by increasing parameters from 355B to 744B and data from 23T to 28.5T tokens.
- Integrates DeepSeek Sparse Attention to cut deployment costs while maintaining long-context abilities.
February 11, 2026 at 2:30 PM
February 11, 2026 at 2:11 PM
New DeepSeek model released. Model available on mobile and free to use for now. Cut off is May 2025, so a new base, 1M tokens claimed. Doesn't support image input yet. Possibly an early checkpoint. 🤔
x.com/i/status/20...
x.com/i/status/20...
February 11, 2026 at 1:58 PM
New DeepSeek model released. Model available on mobile and free to use for now. Cut off is May 2025, so a new base, 1M tokens claimed. Doesn't support image input yet. Possibly an early checkpoint. 🤔
x.com/i/status/20...
x.com/i/status/20...
Sam Altman's Town Hall last month predicts 100x cheaper intelligence by 2027 and a massive shift in software engineering roles.
- Jevons Paradox in coding: Efficiency won't kill engineering but will explode software demand; the role shifts from syntax to "directing" computers.
- Jevons Paradox in coding: Efficiency won't kill engineering but will explode software demand; the role shifts from syntax to "directing" computers.
February 11, 2026 at 12:31 PM
Sam Altman's Town Hall last month predicts 100x cheaper intelligence by 2027 and a massive shift in software engineering roles.
- Jevons Paradox in coding: Efficiency won't kill engineering but will explode software demand; the role shifts from syntax to "directing" computers.
- Jevons Paradox in coding: Efficiency won't kill engineering but will explode software demand; the role shifts from syntax to "directing" computers.