Emilz

Blink and you’re behind. Here's your February AI Recap: • DeepMind Genie 3 • Voxtral Transcribe 2 • Kling 3.0 • Opus 4.6 • Codex 5.3 • Recursive Language Models (RLM) • Prompt to binary • Context rot • Seedance 2.0 • ElevenLabs v3 Expressive Mode • GLM-5 • MiniMax 2.5 • Gemini 3 Deep Think • GPT-5.3-Codex-Spark • Claude PowerPoint • Qwen 3.5 • Grok 4.20 • Sonnet 4.6 • Double prompt (Left to Right) • Cursor Plugins • Phoenix-4 • Gemini 3.1 Pro • Taalas HC1 (17k tokens/s) • Claude Code Security • US-China Distillation Wars • FDM-1 • Mercury 2 • Cursor Demos • Claude Code Remote • MatX One Chip • Perplexity Computer • QuiverAI Arrow 1.0 • Moonlake World Model • Claude Cowork Scheduled Tasks • Nano Banana 2 • Claude Agent Teams • WarClaude • France Gov Spending Data MCP • ClaudeCode Auto-Memory • Tzafon RL (4B) • Apple Xcode 26.3 with Claude/Codex • OpenAI 110B Funding round • while True: Cursor = "dead" Cursor = "back" • FactoryAI Droid Missions • Imbue Evolver • GPT 5.4 Leak • Anthropic - Military Disagreement • OpenAI - Military Agreement • Claude Import memory • US Claude assisted Iran strikes What did I miss? See you in next week's recap.

@0xSero What about the planet man

This is how I end my nights. Today: 1.5 Billion tokens in Codex 22M tokens GLM 51M tokens Kimi 41M tokens Claude 14M tokens MiniMax

@bruvimtired Cmd + shift + P Switch to light mode

uuhhhmm this isnt really glass-y?

@thsottiaux But where

Codex will take us places

When you've used up all your Cursor Cloud Agents free credits, but you're addicted.. bugbot /fast when?

@gabriel1 /prompts: so good, why deprecated

im just copy pasting now with codex 5.4 high (NOT xhigh, NOT medium) before every pr. it's beautiful

@fchollet English business virtuoso fails at easy Full chinese test

This is more evidence that current frontier models remain completely reliant on content-level memorization, as opposed to higher-level generalizable knowledge (such as metalearning knowledge, problem-solving strategies...)

Composer 2 only needs to be almost as good as the previous opus 4.5 model. If the benchmark is real and felt in the code, it's GG. The only thing that made people bearish of Cursor, was the cost of their plans vs the mega subsidized ones from OpenAI and Anthropic. But if they OWN a good enough model, they can start to truly compete with the BIG 2/3 I feel like Auto mode is about to have crazy limits

@cursor_ai The future was foretold

Composer 2 is now available in Cursor.

@IterIntellectus @OfficialLoganK @GoogleAIStudio Don’t worry i’ll release another monthly ai news recap in 10 days

@OfficialLoganK @GoogleAIStudio I CAN'T KEEP UP

Introducing the all new vibe coding experience in @GoogleAIStudio, feating: - One click database support - Sign in with Google support - A new coding agent powered by Antigravity - Multiplayer + backend app support and so much more coming soon! x.com/GoogleAIStudio…

@nikitabier The screen is tired of my face

No, there’s nothing over there. Come back to the screen.

When will we have /fast bugbot man

NEVER EVER TRUST OPUS

Lesson learned. Don't give Opus a second chance at backend.. I need cursor bugbot /fast version ASAP. Waiting 10 min for it's review is way too long

@Tom_Grab @OpenAI They probably are working on a 5/6-o model to satisfy the 4o crowd

@emilheap @OpenAI The best price-to-intelligence ratio is still 4o-mini. Not only price per token is lower, but it’s a very capable non-thinking model so the output tokens are way lower than the new models. I think OpenAI hit its peak in the 4o and o3 era.

GPT-5.4 mini is available today in ChatGPT, Codex, and the API. Optimized for coding, computer use, multimodal understanding, and subagents. And it’s 2x faster than GPT-5 mini. openai.com/index/introduc…

@BrianSOTELI @JamesTakesOnAI @OpenAI I want to believe in AI efficiency self improvement snowball loop. Or just infinite subsidies

@emilheap @JamesTakesOnAI @OpenAI It was previously priced basically at cost, or even a loss. Now it’s being priced to at least break even, or to a profit. It’s preparing to go public. Similar to the 3 year VC funded run of Uber being the best thing on the planet. Cheaper than owning your own car. 🚗

@thsottiaux Tomorrow you’ll release quantum agents. The day after, GPT-AGI-1 will be released which is SoTA in everything. And the day after that we’ll all be sub agents

Yesterday we launched subagents in Codex. Today we released GPT-5.4-Mini, which is SoTA in its category. Coincidence or genius move?

@brianakaka @OpenAI Yeah insane value for mass data processing with good enough reasoning

@emilheap @OpenAI This was my first question. Thanks. imo GPT5mini is best of cheap/fast/smart LLMs, and still has value.

@JamesTakesOnAI @OpenAI Yeah true, though they change purpose the purpose of the mini models as time goes on. The -mini are amazing for large data processing at low cost and good intelligence. But the output price more than doubling kinda destroys that use case

thats the real moat strategy tho. make the best model expensive, then release mini/nano versions that are "good enough" at lower cost. developers get locked into the API ecosystem chasing the flagship benchmarks, then settle for the cheaper tier in production. classic price discrimination playbook

@X_Arcadian @OpenAI Already optimized to the max.. so probably would need to check out specialized open source models that don’t suck at reasoning

@emilheap @OpenAI Yeah I guess if you had a use cases for the old nano for 1 off llm calls that's not going to fly anymore. Wonder if you could male up for it a bit with a smaller prompt amd less Guardrails.

@himbodhisattva @OpenAI It is out yeah

@emilheap @OpenAI probably better to compare 5.4-nano to 5-mini, assuming that comes out