Evi

Earlier this week, we published our technical report on Composer 2. We're sharing additional research on how we train new checkpoints. With real-time RL, we can ship improved versions of the model every five hours.

Today, we introduce Phota Studio and Phota API, powered by our photography model that brings flagship image model capabilities, personalized to you. With personalization, an image model stops being just playful and starts becoming useful for photography. With Phota Studio, you can: - Reimagine composition, lighting, or posture while still looking like yourself - Create editorial, stylized, and studio-quality portraits of yourself, or bring someone you love into the frame - Revive the blurred shot, bring in the person who missed the group photo, fix the awkward expression - all without losing what made the moment worth keeping With Phota API, you can finally build photo experiences where real people are the core. Marketing assets, editorial shoots, wedding photography: workflows that needed identity fidelity that GenAI couldn't deliver. Until now. Ultimately, we want to make compelling photographs accessible to everyone. Phota API and Phota Studio start to make that possible: empowering people to explore, imagine, and create without losing themselves in the image. With Phota Studio and Phota API, developers can build new photo experiences, while photographers and creators can explore a new kind of AI-native editing and generation. The next photo experience starts here!

Codex is already really good and my daily driver. But they really need to fix how terrible GPT-5.4 regardless of reasoning effort is at design/UI. It's seriously really bad. Hopefully this means it gets fixed sooner so I don't have to keep using Gemini 3.1 Pro and Claude.

LLM systems swallow harness progress. The most general/universal LLM innovations migrate from client-side harnesses to server-side tools. Innovation typically happens first inside the harness. For example, AI reasoning was originally a harness around GPT-3 ("let's think step by step"). This approach worked so well that it migrated behind the API as a tool (competitive reasons were also a factor; but general utility dominated). Many wouldn't think of AI reasoning as a tool but it definitely is (it's a tool to do natural language program synthesis -- but that's another topic). The same happened with code interpreter which started out as a client-side harness and moved server-side. These tools are made available at inference time to the model alongside specific training to teach the model when and how to use each tool. Because of this, the line between tool and model can get quite blurry. Best to consider such tools as "internal" to the LLM system. This is actually a good test of how general a harness feature is. If a feature remains "stuck" client-side, say inside codex or claude code, then it's likely very task- or domain- specific. Client-side harnesses typically encode a lot of human G factor for specific domains. Whereas tools, due to usage pressure of frontier LLMs, are required to be as general as possible else they wouldn't make the cut. So if you care about measuring AGI it's a good idea to pay attention to default LLM system capabilities behind high usage LLM APIs. And if you care about bleeding edge research ideas, such as RLMs, it's a good idea to pay attention to harness innovation. Ultimately, AGI will not depend on a harness in the same sense humans don't depend on a harness.

I refuse to believe that most Americans can not read at a 7th grade level If this is “true” then it means we are simply measuring literacy incorrectly; 163M Americans are actively employed Can your 6th grader read an employment contract? Read heavy machinery instructions?

Wow I fucking hate RL research this is genuinely horrible dogshit. Like 99% of my time has been spent figuring out why my lora won't update on my inference server or other technical problems

→ Search the web for live sports scores and stock prices → Find files on my desktop and run shell commands → Write code and solve math problems → Everything Claude Code does — for free The breakthroughs that made this possible: Apple's "LLM in a Flash" paper showed models can page from SSD using unified memory. I proved it works in practice on consumer hardware — not just in a research lab. Google's TurboQuant research showed you can compress KV cache with zero quality loss. I applied this with two server flags and doubled my context window from 32K to 64K tokens. For free. No code changes. The biggest surprise: the 35B model at 2.6 bits per weight was supposed to have "broken" tool calling. Every agent framework I tried failed — infinite loops, no answers. I stopped asking the model to generate JSON function calls. Instead I ask it simple questions. "Is this a search, shell, or chat?" → one word answer. Works perfectly. The tool calling wasn't broken. The protocol was wrong. Both models. Full agent. Same $600 computer: → 35B MoE: 30 tok/s, 2x faster, smarter reasoning → 9B dense: 16 tok/s, 64K context, reads entire codebases I benchmarked everything: → 212 math problems: 86.3% accuracy (3 categories at 100%) → 10 web search categories: 10/10 accurate → Shell commands: finds videos, checks disk space, reads code → MLX vs llama.cpp: tested both, llama.cpp wins for 35B