Michael Joseph Rosenthal

Eng at large. Guy with dog. @software_daily maintainer. Texts of varied structure: Views my own but by all means help yourself.

[Michael Joseph Rosenthal @micimize]

AGI is a bad term and we should seriously stop using it. Sometimes it’s “can make a cup of coffee _and_ solve your taxes” and sometimes it’s “digital god,” and experts in the field use both senses fluidly.

Like if deep sea oil engineers called both oil and water “the wet stuff”

[Michael Joseph Rosenthal @micimize]

@xpasky Could also be innate properties of info processing systems / mind generalization. Ie, any method that can "get there" must avoid certain pitfalls, and the space of possible capable minds is limited

[Michael Joseph Rosenthal @micimize]

@schrockn I've heard this, but tried switching over my workflow and gpt 5.5 seemed really poor at proposal/design doc authorship. Is your XP purely WRT code output?

[Michael Joseph Rosenthal @micimize]

@ThePrimeagen https://t.co/wpbegStYUi has a bit more motivation, but mostly seems to me like anything interesting in this vein is likely covered by @moonbitlang at a far higher level of rigor, for those interested

[Chris Tate @ctatedev]

I built Zero in 3 days.

I didn't expect it to compile.
I didn't expect it to mostly self-host.
I definitely didn't expect it to work at all.

Inspired partly by Bun's rewrite to Rust, Zero started as an experiment. Honestly, the project says more about where AI is today than it does about the language itself.

It took more than 3,000 agent tasks to get here, and it's still nowhere near ready for serious comparisons, benchmarks or evals.

But the goal is bigger than the current result.

The hope is to either create a new language with tooling designed for agents from the ground up, or take learnings and apply it back to existing languages and ecosystems.

The ideas are simple:

1. Make languages (and new versions) easy for agents to learn, adapt to and fix on the fly, even when not in the training data.
2. Build a standard library comprehensive enough that most projects don't need external dependencies.
3. Create a tight, fast development loop that even small models can reliably work with.

I've never wanted to create a programming language.

But after repeatedly running into the same problems, safe but slow builds, fast but unsafe builds, agents struggling with new languages and version changes, wanting faster builds, smaller bundles and better DX, I started wondering:

Could accelerated, agent-driven iteration produce a language and tooling stack designed around these constraints from the start?

So Zero was born.

[Michael Joseph Rosenthal @micimize]

@paularambles The funniest panel is the one of him sipping direct from the cup

[Michael Joseph Rosenthal @micimize]

@MaximeRivest Now now, you wouldn't be trying to _automate_ the intelligence automaton, would you?

[Michael Joseph Rosenthal @micimize]

@timurkuran Maybe in part, but there's good reason to think that's not the lead driver https://t.co/d41bPZSnbA

[Michael Joseph Rosenthal @micimize]

@davidu Flock's security negligence & ongoing handling/response documented by @bennjordan makes treating their impact as an unalloyed good laughable to me. Is there a compelling defense or refutation there?

[Michael Joseph Rosenthal @micimize]

@evanjconrad Now Oracle "oversees" the US joint venture's algo. So, maybe soon we'll also see domestic influence campaigns. I'm sure that'll make it all balance out, in the end

[Michael Joseph Rosenthal @micimize]

@trq212 I cancelled my sub and realized tame opus API usage would be V costly vs the $200 plan. If a user had "you've saved this much vs API" they'd prob be happy vs always wondering if Ant is tricking them.

"But then we can't trick them later" is the cost of trust, nothing short, IMO

[Michael Joseph Rosenthal @micimize]

@trq212 I know from Ant's perspective there's prob frustration with giving people effectively a dollar for 25-50 cents and having them complain, but it's an image issue I think.

[Michael Joseph Rosenthal @micimize]

@alexjcampbell @leeparayno @jarredsumner @trq212 Could be related to https://t.co/VeitInjkT8

[Michael Joseph Rosenthal @micimize]

This kind of tool could afford a decoupled auditing boundary for prompt-injection security-in-depth. If the main agent gets pwnd, the requests might indicate divergence or malice to a gatekeeper.

Such a pattern isn't feasible with classic direct MCPs

[Imbue @imbue_ai]

We built a library so local AI agents can use any tool you have access to: Slack, GitHub, AWS, Linear, Notion, Stripe, self-hosted tools, anything with an HTTP endpoint.

One command, any API, no custom connectors

We call it Latchkey 🔑

[Michael Joseph Rosenthal @micimize]

@ShaulKfir With a simple `docker run` I'm pretty sure things are a lot more straightforwardly secure, but it's all devil in the details. Regardless any boundary is making you way better off than folks out here YOLOcaling it

[Michael Joseph Rosenthal @micimize]

@ShaulKfir My sense is a local container is fairly secure in isolation, esp podman which has a lower attack surface. But for instance, the vscode devcontainer feature auto-mounts some configs and forwards all ports on startup. Unsure how vulnerable that is.

[Michael Joseph Rosenthal @micimize]

@halvarflake @iam_anandv Vscode has good support for remote container dev & devcontainers, but I'd be more wary about ide connections because they tend to do magic stuff to make the XP seamless, which increases surface area.

Do you think local containers can achieve comparable security to remote VMs?

[Michael Joseph Rosenthal @micimize]

@deepfates Another stopgap otw to an outbound visual organ. A few more interface leaps and we'll be close to buffering raw thought into multimedia

[Michael Joseph Rosenthal @micimize]

@ThePrimeagen I guess that means it'll need to provide an API to transmit "age signal" to agents per California's Digital Age Assurance Act AB 1043

[Michael Joseph Rosenthal @micimize]

@ad0rnai Yeah I feel like he misses the point. If bespoke treatments are "trivial" we would prob be focusing on scaling that protocol. "Can't save you, I mean it's trivial to take a 50% shot but we'd need 20y to study that the **right** way. Can't risk killing you before you die"

[Michael Joseph Rosenthal @micimize]

@repligate Is there a particular principle or crux around which you draw this line? Ie do you say this because you're actively working on AI entity continuity and self-empowerment, while much of the ecosystem is oriented around value extraction & precludes discussions of rights warranted?