TSMC bets $31 billion on AI's future
What's going on?
TSMC is dropping $31.4B on three new Taiwan fabs for sub-1.4nm chips with AI data centers now making up two-thirds of its revenue, up from smartphones.
Why it matters?
More advanced silicon means more horsepower for AI. But the first fab won't ship until ~2030, so the compute crunch everyone's feeling isn't easing anytime soon.
How we see this at CometChat
Every AI agent including the ones answering support chats and moderating conversations runs on this silicon. The smarter agents get, the hungrier they are for compute. Worth watching who's actually building the supply.
The AI lead just changed hands
What's going on?
Anthropic is weighing a new model launch after OpenAI's GPT-6 Astra pulled ahead in enterprise spend - days after Anthropic's CEO publicly urged the whole industry to slow down.
Why it matters?
A two-and-a-half-year lead just cracked. When frontier labs sprint, the models under your apps get smarter and cheaper faster than your roadmap can react.
How we see this at CometChat
‘Slow down’ one week, ‘maybe ship’ the next - the tension is real. But if you're building AI agents into your product, model churn is the actual story. Pick an abstraction layer that survives whichever lab wins the month.
Cancer-detecting AI is now free
What's going on?
Alibaba open-sourced Damo Radar - a model that spots nearly 150 abdominal diseases, including cancers, from CT scans, and it beat 23 of 26 radiologists in testing.
Why it matters?
A smaller hospital with no in-house AI team can now run expert-level diagnostics for free. When radiologists used it as an assistant, missed diagnoses dropped 10% and reads got 30% faster.
How we see this at CometChat
The pattern that makes this work - pairing images with clinical reports so the model actually understands context is the same reason AI agents are getting good at conversations. Context is everything. A model that reads the whole picture beats one guessing from fragments, whether it's a CT scan or a support thread.
One config file to rule them all
What's going on?
Claude Code now reads AGENTS.md - the open instruction format already used by Cursor, Codex, and 60,000+ repos so you stop maintaining a separate file for every AI coding tool.
Why it matters?
Fewer redundant configs, fewer symlink hacks. One markdown file tells every agent your build commands, test rules, and code style. Small change, real relief.
How we see this at CometChat
Standards winning over walled gardens is always the good ending. The same instinct applies to how AI agents plug into your product - shared, portable, boring-in-the-best-way conventions beat bespoke glue code that only one tool understands. Interoperability isn't glamorous. It's just what keeps things working.
Voice AI just got 13x cheaper
What's going on?
Google launched Gemini 3.8 Live - voice models that talk, see, switch across 97 languages mid-sentence, and run background tasks while chatting. Extended Thinking topped OpenAI's GPT-Live-1 on quality, at roughly 1/13th the cost.
Why it matters?
Real-time voice agents just went from ‘nice demo, wild bill’ to something you can actually run in production. Price was the wall. That wall's coming down.
How we see this at CometChat
‘Let me check that…’ while the agent works in the background - that's not a model detail, that's conversation design. Voice agents live or die on whether the interaction feels human under real latency. The model's getting cheap. Making it feel good is still the hard part.
Top comments (0)