AI news that matters

1. Hume AI's Audio Decision API analyzes voice without transcripts

   Hume AI released an Audio Decision API that answers plain-language questions about tone, pauses and overlapping speech directly from audio. The company reports 70.7% agreement with human raters and a median latency of about 0.23 seconds.

   15K posts · View on X (opens in a new tab) · Read our story

2. Anthropic's S-1 shows fast revenue growth, large losses and huge compute commitments

   Anthropic's IPO filing reports $33.2 billion in revenue for the first nine months of 2026 and a $104 billion net loss, mostly from non-cash charges, plus about $895 billion in compute commitments. Polymarket odds on the IPO fell after the filing, posts note.

   5.3K posts · View on X (opens in a new tab) · Read our story

3. Debate grows over the limits of vibe coding

   Developers are debating vibe coding, accepting AI-generated code with minimal review. Proponents cite speed for prototypes, while critics point to security risks such as data leaks and say human review stays essential for production software.

   3.4K posts · View on X (opens in a new tab) · Read our story

4. MIT researchers show GPT-6 Astra editing neural network weights directly

   MIT researchers report that GPT-6 Astra can read and edit the weights of smaller networks via a method called Weight-Space Programming. In one test it changed 0.016% of Qwen3.5-0.8B's parameters, lifting math accuracy from 7.9% to 48.1%, they say.

   2.2K posts · View on X (opens in a new tab) · Read our story

5. Developers debate MCP as the standard connector for AI agents

   Builders are debating the Model Context Protocol, the open standard introduced by Anthropic and now under the Linux Foundation. Supporters cite dynamic tool discovery, while skeptics call it REST with extra layers and flag risks like tool-poisoning attacks.

   1.5K posts · View on X (opens in a new tab) · Read our story

6. Gemini 4 Carbon leads Artificial Analysis Intelligence Index

   Google's Gemini 4 Carbon scored 59 on the Artificial Analysis Intelligence Index v5.0, ahead of Claude Opus 5.5 at 58 and GPT-6 Astra at 57. Google is rolling it out first to selected cyber defenders via its Fairwind Program, priced at $2/$10 per million tokens.

   842 posts · View on X (opens in a new tab) · Read our story

7. AI shopping agents stumble on basic purchases in user tests

   A San Francisco user described AI shopping agents failing orders at AutoZone and Sephora because of fraud flags, virtual cards and address mismatches, while one purchase succeeded. Critics including NYT's Mike Isaac question the need for agentic commerce.

   695 posts · View on X (opens in a new tab) · Read our story

8. New tools let AI agents collaborate in shared workspaces

   Designer Winnie introduced RAVEN AI, a coordinated team of agents working toward one goal, and Chi Wang launched S.Space, rooms where different users' agents collaborate. Skeptics point to unresolved permission and reliability issues.

   662 posts · View on X (opens in a new tab) · Read our story

9. Free stealth models Crouton Alpha and Pulse Alpha appear with 1M-token context

   Two anonymous models, Crouton Alpha and Pulse Alpha, launched free on platforms including OpenRouter and OpenCode with 1M-token context and multimodal input. Developers are testing them on coding and agent tasks; the providers have not been disclosed.

   656 posts · View on X (opens in a new tab) · Read our story

10. Google's Harmony Harness lets AI agents rewrite their own tools

    Google's Harmony Harness lets agents revise their own tools, instructions and checks based on feedback, and reportedly beats older methods by more than 10 points on benchmarks like SecRepoBench. Practitioners stress verification loops and human oversight.

    337 posts · View on X (opens in a new tab) · Read our story

Based on public posts on X.