lherron

A few years ago the SRE crowd went through a toil automation phase. SWEs are now gaining the tools to do the same.

The Parallel Anxiety section really stands out for me. It's a new level of FOMO.

Also "LLM Underutilization Anxiety" - I have 15 $20++/mo subscriptions of VC subsidized token generators. If I'm not hitting rate limits, I'm obviously not min-maxing hard enough.

With the Anthropic rug pull on quotas for Max, I feel the short-mid term value sweet spot will be a Frankensteined together “Claude as orchestrator/coder, falling back to local models as quota limits approach” tool suite.

The lack of MCP support in their desktop client is especially disappointing considering everything Anthropic has shipped on Claude Code the last few weeks.

ChatGPT desktop client with only search/fetch MCPs is far, far inferior to CC from a utility/value perspective.

Honestly this article sounds like someone is unhappy that AI isn’t being deployed/developed “the way I feel it should be done”.

Talent changing companies is bad. Companies making money to pay for the next training run is bad. Consumers getting products they want is bad.

In the author’s view, AI should be advanced in a research lab by altruistic researchers and given directly to other altruistic researchers to advance humanity. It definitely shouldn’t be used by us common folk for fun and personal productivity.

Thanks for this! Subscribed.

Even if you don't care about open source, you should care about all the obfuscation happening in the prompts/models being used by Cursor/Claude Code/etc. With everything hidden, you could be paying 200/mo and get served Haiku instead of Sonnet/Opus. Or you could be getting 1k tokens of your code inserted as context instead of 100k to save on inference costs.

Hope this will pressure Anthropic into releasing Claude Code as open source.

Isn't this what sampling is for?

https://modelcontextprotocol.io/docs/concepts/sampling

This is my experience as well. I do not view the first shot LLM generated code as the end result. It’s more like a blob of clay I can continue to shape (with the LLM) until it meets my standards.

I also had a similar typed language experience, switching from untyped to type hinted python made the outputs much easier to understand and assess.

With all the incremental releases, it’s harder to see the advancement. Maybe it would be more fair to compare 4 vs 3 than 4 vs 3.7.

Overly aggressive “let me do one more thing while I’m here” in 3.7 really turned me off as well. Would love a return to 3.5’s adherence.

HN front page isn’t what it used to be.

Repo name checks out.

Looks awesome, will not try!

Factorio blueprints in action.

Feels kinda like JIT compiling your agent prompts into code. Awesome concept, hope it pans out.

I chalk it up to VC subsidized pricing. I use my monthly Cursor quota, then switch to Claude Code when I run out.

I agree for method-level changes, but the more you’re willing to cede control for larger changes, even in a familiar language, the more an LLM accelerates you.

For me, I give Gemini the full context of my repo, tell it the sweeping changes I want to make, and let it do the zero to one planning step. Then I modify (mostly prune) the output and let Cursor get to work.

I love this! I have a similar automation for moving a feature through ideation/requirements/technical design, but I usually dump the result into Cursor for last mile and to save on inference. Seeing the cost analysis is eye opening.

There’s probably also some upside to running the same model multiple times. I find Sonnet will sometimes fail, I’ll roll back and try again with same prompt but clean context, and it will succeed.

Wait till you deal with google genai lib vs google generativeai lib

Agreed! And with all the gaming of the evals going on, I think we're going to be stuck with anecdotal for some time to come.

I do feel (anecdotally) that models are getting better on every major release, but the gains certainly don't seem evenly distributed.

I am hopeful the coming waves of vertical integration/guardrails/grounding applications will move us away from having to hop between models every few weeks.

These one-shot prompts aren't at all how most engineers use these models for coding. In my experience so far, Gemini 2.5 Pro is great at generating code but not so great at instruction following or tool usage, which are key for any iterative coding tasks. Claude is still king for that reason.

Kokoro TTS is pretty good for open source. Worth checking out.

More like: every time you tell o3 to do something, it will first reread the onboarding docs (and charge you for doing so) before it does anything else.

Also this:

https://news.ycombinator.com/item?id=42952605 - Ingesting PDFs and why Gemini 2.0 changes everything

I don't know how you solve the "training data and tooling prompts bias LLM responses towards old frameworks" part of this, but once a new (post-cutoff) framework has been surfaced, LLMs seem quite capable of adapting using in-context learning.

New framework developers need to make sure their documentation is adequate for a model to use it when the docs are injected into the context.

ollama is making this worse by not denoting what you’re getting in the model name. You have to look at individual model cards to see it’s distilled.

It’s a question of depreciation of the GPU vs time to deploy AI.

If RL and synthetic data creation are all we need for self-improving AI, the large labs have an internal use case to utilize all available compute, regardless of efficiency gains.

If they have to go external and deploy in the broader market, time to deploy would make the GPUs worthless before they are fully utilized.

You don’t know how to use the three seashells? lol

I do both. I have a basic template that also includes my specific AI instructions and conventions that are inputs for Aider/Cursor. Best of both worlds.