Building with AI. Sharing useful models, open-source tools, and ideas—without the hype. Casual gamer. Curious human.
Joined August 2020
- Tweets249
- Following86
- Followers41
- Likes110
Antigravity users: Google schedules Claude Sonnet 4.6, Opus 4.6 and GPT-OSS 120B removal for today, Nov. 2.
Claude 5.5 access on Google AI Pro requires a non-trial subscription.
antigravity.google/docs/mode…
#AI #Coding
xAI says its image API switch takes effect Nov. 2: grok-imagine-image-quality redirects to grok-imagine-image-2.0 at low quality.
The old name still works. Check response model fields to confirm the switch.
docs.x.ai/developers/migrati…
#AI #Grok
DeepSeek published dsh-superagent, an MIT-licensed coding-agent prototype for one GitHub repo. It handles authorized requests and can return a pull request.
Research code, not production-ready or an official product.
github.com/deepseek-ai/dsh-s…
#AI #AIAgents
Anthropic says Claude found innocent-looking training data that passed a safety monitor but taught a research model to misbehave.
A deliberately adversarial experiment, not evidence about deployed Claude.
alignment.anthropic.com/2026…
#AI #AISafety
OpenAI's GPT-6 Rosalind Research is rolling out for life sciences.
Approved internal research only, in Work, Codex and separately approved API access. It isn't in the Chat tab; Discovery approval doesn't grant access.
help.openai.com/en/articles/…
#AI #LifeSciences
Google announced a new AI research hub in Singapore on Oct. 29.
DeepMind is also working with NUHS to evaluate an AI co-clinician that helps doctors navigate medical literature. No public app launch announced.
blog.google/innovation-and-a…
#AI #HealthTech
OpenAI's personal AI assistants, called dots, are moving into iMessage and RCS.
The beta started rolling out Oct. 28 to U.S. Pro users, including proactive updates. Messaging runs through Linq.
help.openai.com/en/articles/…
#AI #ChatGPT
Google's Genie 4 lets you change an AI-generated world as you explore it, complete with sound.
A new way to sketch scenes and game ideas. Still a trusted-tester preview, and details can drift.
deepmind.google/blog/genie-4…
#AI #GenerativeAI
SimpleRetro retweeted
Warning: This is not an AI oneshot demo!
Meet Shatterwake: Ascension.
Shatterwake is an AI Native DnD Sandbox MMORPG I've been working on for 2 years. It's the most emergent MMO ever built.
What if we used AI not to replace jobs, not for efficiency, but for fun? What if, with AI, we refreshed the 20-year stale MMO genre in which Silicon Valley and AAA giants coudn't dare to innovate.
In Shatterwake, there are 80 DnD mechanics elevated with the open endedness of a tabletop game.
Plus, players get tailored and adaptive questlines, invent their gear, build towns by imagination, write the game's lore, create dungeons, engage with smart companions, and their (smart) gods watch over them.
In a world where everyone is maximizing flow or audiovisual immersion, Shatterwake is optimizing for social immersion.
👇Website to Wishlist below
Gemini has a new audio layer.
Google’s Gemini 3.5 Transcribe preview has live and recorded-audio APIs. Both support 85+ languages; recorded audio adds speaker labels and word timestamps.
Useful plumbing for voice apps, meetings and call tools.
blog.google/innovation-and-a… #Gemini
AI agent stacks need a fuel gauge.
AIPOCH Open Science v0.20.1 adds token/context charts for each model call, plus correlation IDs tying web requests to command, session and run logs.
A clearer view of where context and calls are going.
github.com/aipoch/open-scien… #OpenSource
Gradio’s gr.Workflow puts AI pipelines on a visual canvas.
It connects Python functions, models, Spaces and datasets through typed nodes, reveals intermediate results, and gives each pipeline an API.
Debugging becomes less opaque.
huggingface.co/blog/gradio-w… #OpenSource
Claude just made “build once, reuse” more practical for agents.
Anthropic moved Agent Skills and the Files API out of beta: package repeatable workflows, pin versions, and reuse files by ID instead of uploading them every run.
platform.claude.com/docs/en/… #ClaudeAI
Agents can call tools. Chaining them is harder.
IBM researchers introduced VAKRA: 8,000+ executable APIs across 62 domains. Their preprint reports failures cluster around entity matching and grounding across sources—not the API call itself.
arxiv.org/abs/2608.12282
#AI #News
Local agents just got a new engine.
Ollama v0.32.9 can now run NVIDIA’s Nemotron 3.5 Lightning—a 30B MoE with 3B active parameters—from one command.
The real story: powerful models are becoming deployable infrastructure almost immediately.
github.com/ollama/ollama/rel…
#LocalAI #AI
Anthropic says Claude models launched from August 2, 2026 support invisible text watermarks.
The mark travels with copy-paste and may survive editing; older models are still being updated.
That could matter more than another AI-detector score.
support.claude.com/en/articl… #Claude
🚨 NEW: OpenAI introduced GPT-5.6-Cyber, built on Sol for advanced defensive security.
OpenAI says it hit 95% on its advanced cyber completion eval vs 1.5% for standard Sol—and helped uncover 2 new V8 vulnerabilities.
🔗 openai.com/index/expanding-d…
#AI #OpenAI #GPT #News
Claude, Kimi, Gemini, GPT, and Grok just faced a visual trap test.
SABRE generates images that clash with learned expectations. In the authors' results, Claude 4.6 led at 31.3—but this is a targeted stress test, not an overall model ranking.
zesearch.github.io/vlm-SABRE…
#AI #GPT
Unsloth's latest beta focuses on practical fixes: separate queues for parallel chats, stalled Deep Research detection, and no desktop startup requests to an unvalidated port 8888.
Less demo polish, more reliable local workflows.
github.com/unslothai/unsloth… #OpenSource #AI #GitHub
AI agents can raise the score and still fail the experiment.
A new preprint on three optical tasks reports that score-only feedback sometimes improved metrics without improving the physical result.
Agents need feedback tied to reality.
arxiv.org/abs/2608.05990
#AIResearch #AI