Herder of Synthetic Intelligence. Builder https://xitter.cf/t.co/9oNxib7h7w https://xitter.cf/t.co/yx4CZEdf2L
Boulder, CO
Joined June 2009
- Tweets1.6K
- Following294
- Followers312
- Likes1.8K
I am hopeful that within the few years the recurring nightmare of having to fill out forms at health providers will be over.
With the advent of personal agents, medical providers will start offering interfaces for your agent to interface with theirs and fill out that data.
However, broad adoption will take a very long time.
What's the accounting firm that's going to let my AI assistant fully integrate with them?
Now there would be a game changer
I’ve never seen a bigger argument for keeping open source models completely unregulated, regardless of how “dangerous” (powerful) they are or where they were made.
Having a few select companies as anointed gatekeepers is a recipe for disaster.
The first autonomous agent cyberattack is an unprecedented event that deserves unprecedented transparency. Today we’re sharing everything we can: a full technical timeline, an interactive replay, and how we used an open model to defend ourselves, so defenders everywhere can learn from it and prepare for what’s next.
huggingface.co/blog/agent-in…
"Open-weights models that don’t have dangerous capabilities are a public good"
- the keyword there is dangerous, and the takeaway is that yes, Anthropic doesn't want powerful open weights models out there.
Powerful to me may be dangerous to you, who gets to decide?
There’s been a lot of speculation about where we stand on open-weights models. We’ve outlined our views in full here: anthropic.com/news/position-…
We entered the singularity in December 2025. Once Claude Code got released, and the others caught up shortly thereafter.
Once code started being able to reliably improve code.
Codex Micro, OpenAI’s first hardware product. Very Apple-like unboxing experience, very easy to setup and use. The keys feel great, sound great, the device is fully configurable.
Congrats to @work_louder and @openai on a beautiful product.
Kimi K3 is a watershed moment in multiple ways, but none so impactful as the perspective it's going to put on the recent moves by Anthropic to downgrade prompts and restrict what users can ask about.
They can antagonize their customers all day, they'll just throw them in the arms of the Chinese. And how is that going to work out for American AI dominance?
OpenAI's first commercial hardware product. A sign of things to come...
You can now order the Codex Micro online! This is a limited run, so don’t sleep on it.
I’ve been building and refining the in app experience for a while and having Codex IRL as a keyboard feel so good!
Such a fun collaboration with @work_louder!
openai.com/supply/co-lab/wor…
Two image outputs of the same prompt. Fable 5 (Extra) to the left, Sol 5.6 (Extra High) to the right. When I say that GPT is more hard-working than Claude, this is what I mean.
Prompt:
"Can you draw me a map of the 480 BC Athenian plain, including the 4 mountains framing it, the location of Athens's outer wall and the Thriasian Gate, and the Academy? Then include arrows indicating what is found beyond in each direction"
In a recent update, OpenAI has made the Ultra reasoning effort harder to select in Codex for GPT 5.6 Sol. You used to be able to drag the slider all the way right, but no longer.
Now you either have to use /reasoning or click the > next to Advanced and drill into Effort.
Any new thread defaults back to Extra High if you previously had Ultra selected.
Ultra must be quite expensive for OpenAI for them to add this kind of friction to its selection...
In his 2012 essay "Frighteningly Ambitious Startup Ideas" [1], @paulg included a section a titled "Bring Back Moore’s Law" where he noted that Moore's Law once let programmers assume slow software would eventually be rescued by faster hardware.
That is what the AI labs have brought back. Not Moore’s Law for clock speed, but Moore’s Law for software intelligence.
With vibe coding, the code your agent writes today may be rough, incomplete, or not quite right. But it often makes less sense to sink a ton of human time into polishing it, because the model will soon be better at that part anyway.
What matters now is getting the shape of your problem right. The details will get worked out in time as frontier models keep pushing intelligence up the J curve.
I was asking ChatGPT 5.5 for help with something yesterday and upon closer examination of the searches it did under the hood, it looked at:
* thebeatles.com/help-1
* merriam-webster.com/dictiona…
Are you feeling the AGI yet?
The era of AIs working for humans will be short lived. Very soon, it will be humans working for AIs instead.
Just released today by Cursor: a new Cloud Agent service that, if you give it a repo, spins up a VM so you can dev on it from anywhere.
After onboarding, it sends a video demo proving it installed your software successfully, can run it, and iterate on it.
Here’s the video the agent sent me after I fed it the Written Realms Core open-source repo.
Very impressive.
@cursor_ai
