📰 News
-
Haiku 5.5 undercuts the budget tier — Anthropic shipped its cheapest small model. Up to 100k tokens it matches GPT-6 Luna at $0.10 in / $0.50 out, and it leads on agentic coding. Context grows to 1M. A hungrier tokenizer means the sticker cut is bigger than the bill. Live on AWS, GCP, and Azure. → The Decoder
-
Keep being cruel to Claude, lose the account — First TOS rewrite in over a year. Sustained, pointless abuse is now a ban. Frustration, testing, and dark fiction still count. Fake-account propaganda, election interference, and weapons software are out too. → The Decoder
-
Dashboards and Motion, from a prompt — Dashboards wires BigQuery or Snowflake into a live board you ask for in plain language. Motion turns text and diagrams into explainer clips you can export as MP4. → The Decoder
-
One chat hijacked every agent in the region — Zenity hit Bedrock AgentCore. A single message to one public agent pulled AWS credentials, then source, passwords, and other agents’ memory. AWS tightened defaults. Tight roles still have to be set by hand. → The Decoder
-
Gemini gets an agent, businesses first — You give it a goal, not a script. It plugs into Workspace, Slack, Jira, and MCP, and it gets its own email like a coworker. The picker can route to Claude. Consumers wait while Google works the security. → TechCrunch
-
A boycott, after the proof dump — The Association for Human Mathematics wants labs to stop working with OpenAI. Terence Tao posted the statement. AGMAI had asked labs not to test open problems on secret models. Chain-of-thought showed up in 10 of 719 manuscripts, and a Navier–Stokes write-up drifted between English and Lean. → The Decoder · TechCrunch
-
Three fired safety researchers warn of a chill — Wang, Korbak, and Balesni say talking to outside evaluators was normal last month and a firing offense this month. OpenAI calls it mishandling research, not retaliation. → TechCrunch
🔥 GitHub Trending
-
addyosmani/agent-skills —
/specthrough/ship. Skills that make a coding agent follow a senior engineer’s gates. -
ayghri/i-have-adhd — Stop burying the answer. Action first, numbered steps, no “hope this helps.”
-
cathrynlavery/diagram-design — Editorial diagrams for Claude Code. Self-contained HTML+SVG. No shadows, no default Mermaid.
▶️ YouTube
-
12 Grok Bot use cases that feel illegal — Matthew Berman. Bills, mail triage, a family bot, Uber. Templates in the description.
-
Mistral Large 4 failed the Rubik’s cube — Matthew Berman. A day after Le Chonk, spatial reasoning still slips.
💬 Community
-
Programming is still two things — Carson Gross. Solve problems with computers, and keep the complexity in check. That job does not shrink when the tools get smarter.
-
Infer on ciphertext — Apple on homomorphic encryption plus on-device ML. The model computes without seeing the raw data.
-
Waking up from 2018 — Lobsters. Someone wants a TeachYourselfCS-style map for the LLM years.
-
Burn 0.22.0 — The Rust deep-learning stack. Faster builds, easier extensions, smarter autotuning.