mbil

Agree. I think they're intentionally sitting on the fence between "These models are the most useful" and "These models are the most dangerous".

They want the public and, in turn, regulators to fear the potential of AI so that those regulators will write laws limiting AI development. The laws would be crafted with input from the incumbents to enshrine/protect their moat. I believe they're angling for regulatory capture.

On the other hand, the models have to seem amazingly useful so that they're made out to be worth those risks and the fantastic investment they require.

At the end of last year I took the old VCR and VHS home movies from the 90s out of my mom's basement and digitized them using this cheap converter[0]. I also had to get some rubbing alcohol and foam swabs to clean the VCR heads, which only took a few minutes[1] (some of the older tapes would play normally for a moment before the quality degraded into poor tracking and static). My family enjoyed watching them.

[0]: https://www.amazon.com/dp/B0CB6BW6X4

[1]: https://www.youtube.com/watch?v=5ClXrffkN3M

The PubPeer discussion of the original paper is kind of interesting https://pubpeer.com/publications/A32D7989007655CBF8D9DB2A250...

Well it's mentioned in the tweet this thread is about

> Anyway there are many others - e.g. nanobot, zeroclaw, ironclaw, picoclaw (lol @ prefixes).

It's possible to use opencode with the plus subscription using this plugin for auth [0][1]. Just tested this and it appears to work.

[0]: https://opencode.ai/docs/ecosystem/#:~:text=Use%20your%20Cha...

[1]: https://github.com/numman-ali/opencode-openai-codex-auth

Reminder that you can press the iPhone power button five times to require passcode for the next unlock.

> I've written tons of code, I've taken no shortcuts, and I've focused on improving over my many iterations. This has enabled me to be an effective steward of generative coding (etc) models, but will younger engineers ever get the reps necessary to get where I am? Are there other ways to get this knowledge and taste?

I think this will be a problem in the middle term, and I've written about such deskilling before [0]. With the latest crop of foundational coding models and harnesses, and more progress on the way, I'm beginning to wonder if it will matter? If there's a future where agents are designing the code, implementing the code, and reading and reviewing the code... At that point the code is no longer the thing. "Software engineers" will continue to sit at the interface of product and software, but the software will be writing itself. Of course there will be a need for programmers who can actually read and write computer code, the same way there's a need for Fortran and compiler devs today.

The skill that all software engineers will need to learn, regardless of level, is how to leverage commoditized reasoning to build products effectively.

- how to design systems declaratively and in terms of requirements and constraints

- how to configure the systems in such a way that they're automatically testable end-to-end

- how to move tacit knowledge out of people's heads and into the context (all of our meetings will be transcribed; questions from the agent will be generated during the meeting resolve ambiguity; the agent will be an omnipresent attendee in all meetings: "Agent: The topic you're discussing overlaps with what Sally said three days ago when she met with Mike. They covered xyz..."; companies that follow remote work best practices will have an advantage here)

- how to allocate and orchestrate teams of people and agents

[0]: https://matthewbilyeu.com/blog/2025-03-08/ai

thanks for sharing this, it resonates with me. as some other commenters have said, CLIs, cURL, and other shell tools are composable and discoverable. it seems like a good direction.

auth considerations are present in the design of MCP. this, as opposed to the hodgepodge auth story with CLIs. there are APIs that either don't support OAuth or where using bare credentials is more expedient, and using agent-visible env vars is a security incident waiting to happen. but that doesn't necessarily mean we must use MCP. i think it's a matter of time before agentic tools come bundled with a proxy layer from which secrets / env vars can be set and used but not directly read [0].

[0] https://www.joinformal.com/blog/using-proxies-to-hide-secret...

The namesake for this emacs minor mode for writing: https://github.com/rnkn/olivetti

there's also https://github.com/cdarwin/go-koans for small go exercises.

you may have issues running in Docker; when i last touched this i needed to modify docker.sh:

  -docker run --rm -ti -v "$PWD":/usr/src/koans -w /usr/src/koans golang:1.6.0-alpine go test
  +docker run --rm -ti -e GO111MODULE=off -v "$PWD":/usr/src/koans -w /usr/src/koans golang:1.18-alpine go test

Doesn't seem to agree with the data on https://layoffs.fyi/

I'm using mcp-agent and have tried the orchestrator workflow pattern[0]. For deep research I'm having mixed results. As far as I can tell, it's not using prompt caching[1] with Anthropic models, nor the gpt-5 responses API[2], which is preferable to the completions API. The many MCP tools from a handful of servers eat up a lot of context. It doesn't report progress, so it'll just spin for minutes at a time without meaningful indication. Mostly it has been high cost and high latency without great grounding in source facts. I like the interface overall, but some of the patterns and examples were convoluted. I'm aware that mcp-agent is being worked on, and I look forward to improvements.

[0]: https://docs.mcp-agent.com/workflows/orchestrator

[1]: https://docs.anthropic.com/en/docs/build-with-claude/prompt-...

[2]: https://platform.openai.com/docs/guides/migrate-to-responses

I made a musical game of life not long ago if you want something to play with https://matthewbilyeu.com/tone-of-life

When I was a kid, my dad and I were watching a cooking show together. I asked him "what do they do with all the food they make", and then, as if on cue, the host said, "In case you're wondering, the staff eat all the food we make here." My dad and I looked at each other with a silent look of "whoa".

From the title I expected this would be like talk radio (like NotebookLM style) discussion of random wiki pages.

Thanks for sharing. I’m a “destination” programmer most of the time, and I’ve welcomed LLMs into my work. I actually wrote about this from my perspective just the other day: https://matthewbilyeu.com/blog/2025-06-14/vibecoding-s-allur...

This 2012 study in humans says

> Since this was also a noninterventional study, two possibilities for interpretation of the findings were either that a high cysteine somehow promotes obesity or that obesity influences cysteine turnover, thereby raising plasma tCys.

https://onlinelibrary.wiley.com/doi/full/10.1038/oby.2011.93

So soon there will be programs to generate papers keystroke by keystroke to mimic human drafting. It’s an arms race

I think we’ll see a lot more software. Lots of non tech people will increasingly have the ability to create custom software tools and prototypes. They’ll share and remix these. Some of them will be fully productized, at which point professional software engineers will be called in to help untangle the LLM spaghetti and create proper applications. So I think there will be more software work in consultancy.

I think we’ll see advancements in robotics and more hires there.

And I think there will be more jobs around the LLM ecosystem — progress on foundational models, inference optimizations, on prem migrations, networks of agents, AI more deeply integrated with existing sw.

Overall I think there will be more jobs in observability, security, and infrastructure.

I agree there will be fewer junior positions. I’ve written about some of these ideas before including a deskilling for new practitioners https://matthewbilyeu.com/blog/2025-03-08/ai

I agree that there’s a lot of ambiguity and tacit information that goes into building code. I wonder if that won’t change directly as a result of wanting to get more value out of agentic AI coders.

> Sure, if you write down a summary of each and every meeting you've attended for the past 12 months, as well as attach your entire company confluence, into the prompt, perhaps then the LLM can design the right architecture. But is that realistic?

I think it is definitely realistic. Zoom and Confluence already have AI integrations. To me it doesn’t seem long before these tools and more become more deeply MCPified, with their data and interfaces made available to the next generation of AI coders. “I’m going to implement function X with this specific approach based on your conversation with Bob last week.”

It strikes me that remote first companies may be at an advantage here as they’re already likely to have written artifacts of decisions and conversations, which can then provide more context to AI assistants.

What level does that make this comment?

Nice setup. I use Jekyll but recently moved to a flow where I author posts as rich text emails and a process on the receiving end persists them as markdown and compresses images. It could be improved but so far has been working well.

https://matthewbilyeu.com/blog/2025-04-06/blogging-via-email

Nice idea! I think it would be useful if the jury members could provide context on their decision as well as see the other jurors’ responses.

Humans aren’t the largest biomass

I’ve been thinking lately that email is a good interface for certain modes of AI assistant interaction, namely “research” tasks that are asynchronous and take a relatively long time. Email is universal, asynchronous, uses open standards, supports structured metadata, etc.

Dr Casey Means has a thought provoking book on this subject Good Energy. It’s about how cell metabolic dysfunction could be implicated in a wide range of health conditions.

I’m an LLM user but I haven’t looked into plugins before. It doesn’t look like they use MCP under the hood, though I’d guess they could?

> It’s almost kind of more impressive that the Incan empire of maybe 10 million people got by on the barter system.

In _Debt_, Graeber argues that it is largely a myth that pre modern civilizations relied on barter, and instead used different systems for tracking debt between individuals and entities.

He says this of the Incas knotted strings in a footnote:

> I note in passing that a study of the Inca khipu system itself would itself be quite fascinating in this regard; the strings were used to record both obligations we would consider financial, and others we would consider ritual, since as in so many Eurasian languages, the words “debt” and “sin” were the same in Quechua as well

I wrote about this recently. I agree, though I think it might take a little longer. There will be a deskilling where the easy problems are solved by AI, so new grad workers won’t have gradual experience to grow into seasoned experts.

I agree with you. I recently wrote up my perspective here: https://news.ycombinator.com/item?id=43308912