Legendaryry

google just quietly became the only company that can generate text, images, video, AND music inside one app.
lyria 3 dropped today. here's why that matters more than people think.
the multimodal scoreboard right now:
Google Gemini: text ✅ images ✅ video ✅ music ✅
OpenAI: text ✅ images ✅ video ✅ music ❌ (coming)
Meta: text ✅ images ✅ video ✅ music ❌
Anthropic: text ✅ images ❌ video ❌ music ❌
google just checked every box first.
lyria 3 does text to tracks, image to tracks, and video to tracks. 30 second songs with custom lyrics. you upload a photo of your dog and it writes a song about your dog

Anthropic: Here’s an Agent SDK but you’re not allowed to use it.

my agent kept writing configs that broke everything then trying to fix what it broke by breaking more things
i’d have to walk to my MacBook, figure out what it changed, have codex fix it and restart it
so i built DocClaw. it indexes the actual OpenClaw docs so the agent follows the right patterns instead of freestyling
less babysitting means more time to build things

big tech companies are burning $500B+ on AI data centers
Apple said nah. invested early in their own silicon. put 2 billion devices in people’s hands. now they can just sit back while everyone else bleeds cash
smartest AI strategy might be not playing the AI spending game
(siri still sucks though)
Discussed this on the @modernmarket_ today

So deepseek v4 today? 👀

468M tokens on codex. one monday.
7h on gigabrain memory integration
3h auditing openclaw health (found 2 sneaky bugs)
5h reviving viral machine ops
3h wiring an AI coaching pipeline
i am completely cooked but everything passed

Just filed my first 3 upstream OpenClaw bug reports on GitHub. feels cool somehow

Why does codex feel more stupid in the app than in the terminal?
Codex app with full permission can’t fix OpenClaw configuration issues that codex tui has no issue handling
Both had reasoning set to high

DeepSeek V4 leaked:
HumanEval: ~90%
(community leaks on Reddit + Skywork AI summaries)
SWE-bench: “>80%” territory
(referenced in Wavespeed comparison posts extrapolating from Claude Opus benchmarks)
Context: up to 1M tokens
(rumor package repeated in Reddit AI Insider + blog recaps)
Cost: allegedly 20–40x cheaper than OpenAI
(based on comparisons to DeepSeek V3/R1 API pricing vs frontier model tiers)
Expected: Feb 17 (Lunar New Year timing widely reported in leak writeups)

china just dropped a wave of frontier AI models in one week. the US is falling behind.
this is february 2026:
USA 🇺🇸
Opus 4.6 (shipped Feb 5)
GPT-5.3-Codex (shipped Feb 5)
Gemini Deep Think (not a new model but strongest reasoning capabilities)
GPT-5.3 general (Polymarket: ~49% by Feb 28)
Sonnet 5 “Fennec” (leaked. imminent.)
China 🇨🇳
GLM-5 (Feb 11. 744B params. open weights. highest open-weight Intelligence Index score yet)
MiniMax M2.5 (Feb 12. SWE-bench 80.2%. ~10% the cost of Opus per verified task)
Doubao 2.0 (Feb 14. ByteDance says GPT-5.2 level at ~10x less)
Qwen 3.5 (imminent. co

i wish uber had a setting to exclude prius drivers