(1/10) Today, we’re sharing new research from @OpenAI: How does CoT monitorability (no-CoT capability and CoT controllability) evolve over the model training lifecycle?
We emerged with unsurprising but sobering findings: as we scale toward more intelligent systems, models are becoming harder to monitor through their CoT. 🧵
Today, we’re sharing the training infrastructure behind Reve models.
In the past two years, we built and scaled our platform to support frontier-level training efforts. We're releasing this technical report as a helpful reference for other small teams building at scale.
reve.com/training-infra
Opus 5.5 is so powerful it made my previous workflow for making teaser videos for papers seem over-optimized, which consisted of an iterative loop of storyboarding and rendering.
I simply threw my paper PDFs at it, asked it to generate @3blue1brown-style animations, and let it crunch for an hour. Not only did it generate perfectly coherent animations that adhered to the 3b1b style, it also came up with a lot of nice visualizations and toy experiments that illustrated core concept completely unprompted! Zero hand-holding.
To celebrate entering the last year of my Ph.D., here are the one-shotted animations for some of my selected past projects, starting with Foveated Diffusion (bchao1.github.io/foveated-di…).
The moving gaze visualization at 0:16 and the pipeline and 0:24 is so cool!
Genomes are full of dark matter of unknown functions.
We present Minerva, a new method for discovery guided by genome language models. Minerva reveals the interactions hidden in non-coding DNA, pointing to hundreds of new putative RNAs and repetitive elements per bacterial genome. Minerva allows us to find and study elements invisible to traditional methods at orders of magnitude greater scale than before.
New essay for @every, by @afterimage
We believe we will look back on Astra as a turning point in how creatives use AI in video work.
For many months we were experimenting with different AI video approaches: generative workflows in ComfyUI, procedural animation with Python, motion design through the Cavalry and After Effects MCPs, and more.
All of these were fundamentally limited by a lack of control. Generative models could sometimes make cool images, but changing one thing without disturbing everything else was difficult.
Many of these generative tools weren't (and still aren't) built for true professional workflows.
On the other hand, general models starting around Opus 4.5 were decent at writing python code to deterministically edit images and videos, but the models were really bad at “seeing” the images and videos they were supposed to be editing, which made them impractical.
The day Astra launched, just from the rumors and marketing, our internal sense was "this might be the one". Then, when we saw the level of 3D detail people were able to achieve with Astra in Blender, we really suspected the model was finally able to "see" and iterate toward things that actually looked good.
While everyone raced to build cool (but often useless) demos of 3D worlds, we got interested in something else: compositing 3D assets into live action footage at commercial-grade quality. To this day, we are quite surprised by how few Astra demos focus on this ability. It feels like a breakthrough for us.
We tested a range of processes, from traditional VFX skills like masking and tracking to hybrid approaches like compositing Blender renders and generative imagery with real footage.
At the highest levels, Astra doesn’t replace someone with true domain expertise in VFX artistry. But, through these tests, it started to feel for the first time that the LLM was actually a viable tool in the kit. It worked well in concert with other tools to make professional results we could actually see ourselves using.
The coolest part about all this is that it makes it easy and painless to attempt visual effects on even the smallest projects and most off the cuff videos. We’re excited about the combination of casual footage and visual maximalism. Effects can now be completely over the top without feeling like the whole reason a shot exists.
Full "Vibe Check" with video examples & side by side Astra vs Sol comparisons (gift link) ➡️ every.to/vibe-check/vibe-che…
Astra x GPT-Images 2.5!
Combining Astra's agentic perception with Image 2.5's consistency leads to some super fun results!
Astra can create, edit, and animate coarse guidance, then render it with GPT-Image
🧵
Some ideas are easier to draw than describe.
Use Sketch to draw right in ChatGPT and show it exactly what you have in mind.
Just type “@ Sketch” in ChatGPT.
Today, we're introducing @Intelligence_ai.
In 6 months, as a team of 10, we scaled from $5M to $60M ARR and 5.5M users across 190+ countries.
We raised a $7.9M seed, led by @IndexVentures with participation from @conviction, @A_StarVC, and @combinator to build DesignArena, a universal interface for accessing and evaluating the world's AI capabilities.
Most evaluations try to simulate the real-world. We believe the real-world is the ultimate verifier.
People come to @DesignArena with a request. Models compete to fulfill the request, and users determine what works best for them.
Their live user behavior evaluates the models, improves how work is routed, and helps people access the right intelligence.
We've helped the world's leading frontier labs break the news on their SOTA capabilities.
What's the limit? Join us and find out.
The most startling creations happen in the messy part, in the iterations, in the edits.
Reve is built for this in-between. Now, images are malleable. Our creative choices are the forces that shape them.
Today, Reve is reinventing templates.
Traditional templates have been severely constrained by primitive, layer-based engines, resulting in overly simplistic, generic, and obviously composited assets. Reve Templates, on the other hand, are infinitely customizable and perfectly harmonized, meaning every asset looks entirely bespoke. And you can use any product image — even photos taken on your phone.
We're launching with about 150 templates, but there are thousands more coming. You can even turn any image in the world into your own template, which you can share with your team.
Reve Templates are live right now. Just click and create.