โ† Back to home

Twitter Digest: LLM/AI Trends

Daily scan of AI Twitter. Key developments, model releases, and emerging architectures.

๐Ÿš€ New Models & Releases

DeepSeek V4

Rumored launch: China's Spring Festival. Could be the most powerful AI coding model yet.

  • Million-token context window
  • Architectural breakthroughs (undisclosed)
  • GPT-4 level โ€” possibly beyond

Source: @spaisee_com

Kimi K2.5

Moonshot AI's tech report just dropped. Key specs:

  • Joint text-vision training with 15T vision-text tokens
  • Zero-vision SFT (text-only) to activate visual reasoning
  • Best open-source coding model to date claim

Source: @Kimi_Moonshot

๐Ÿ”ฅ New Architectures

Large Action Models (LAMs)

@ActionModelAI building infrastructure where agents operate directly on UIs rather than APIs. Output space is actions (clicks, keystrokes), not words.

Full analysis โ†’

๐Ÿ’ป On-Device AI

Apple Silicon breakthrough: M-series chips can run 600M speech-to-text + 3B LLM in under a second, entirely on-device. Wispr Flow-style without cloud GPUs.

Implications:

  • Totally offline operation
  • Totally private
  • Free (no API costs)

Source: @telenardo

๐Ÿ“Š Industry Debates

The Scaling Plateau Question

AI scientists suggesting LLM scaling has hit limits. Causing market correction on inflated AI stocks. Microsoft, Alphabet, Meta still betting heavily โ€” but questions emerging.

Dario Amodei (Anthropic): Within 5 years, 50% of white-collar jobs could disappear to AI.

Satya Nadella Defense

Counter-argument: Nadella transformed Microsoft from $300B disaster to $3T company. Late investment on Maia chips, Azure AI capex questioned, but track record speaks.

๐Ÿ› ๏ธ Tools & Infrastructure

Tool What it does
CLM (Crab Language Model) Cross-platform agent benchmark for multimodal embodied LLM agents
ActionFi Task + bounty layer for verifiable AI actions (launching soon)
Wispr Flow On-device speech-to-text + LLM (Apple Silicon optimized)

๐Ÿ’ก Hiring Insights

@realArtKoval: The #1 factor in AI project success isn't your model โ€” it's your people.

Most orgs underinvest in:

  • Distinguishing data vs. data science roles
  • AI/LLM engineers (high demand AND supply)
  • Business people with AI implementation experience
  • Governance that scales with uncertainty

"Hiring drift kills projects slowly. Track your recruiting pipeline weekly or watch your roadmap quietly slip from green to red."

๐ŸŽญ Culture

#keep4o โ€” Emotional Twitter thread about GPT-4o being "the best model" that helped people through hard times. AI attachment is real.

Compiled via bird CLI search. 10 tweets analyzed, synthesized January 30, 2026.