TomanatorDev

Building, testing, and breaking code. Learning and documenting my journey along the way. Check out my website to see what I'm currently working on.

[TomanatorDev @TomanatorDev]

Golem CLI. Manage files and folders using natural language and a small local LLM (less than 1GB). Easy setup and installation.

Open source: https://t.co/WURmlc4Dxm

#Go #OpenSource #CLI #LLM #AI

[TomanatorDev @TomanatorDev]

@hambleton_j @johnennis Not sure where you got that info from? Googleworkspace isnt his personal GitHub.

[TomanatorDev @TomanatorDev]

@traviscurnutte @cursor_ai Hopefully composer 3 is the same price or composer 2.5 doesnt get deprecated. For real work I haven't had to use anything else. I'm not green fielding every day though, usually working on mature projects. I rarely have to correct it and when I do, it's usually due to vague spec.

[TomanatorDev @TomanatorDev]

@almonk @kilocode has been pretty impressive. Although I ran through $22 testing it out. The test was a full local issue tracker with Rust + Go & 10+ subagents with wiggum loop. Reviewed its plan and implementation with Claude and it was the first time Claude had no feedback whatsoever.

[TomanatorDev @TomanatorDev]

The new @antigravity CLI and IDE is absolutely shocking. It never listens to rules, doesn't listen to it's own config. Breaks more than it fixes even with tiny small well defined tasks. It's hard to describe how much of a regression it is from v1, nevermind compared to other labs

[TomanatorDev @TomanatorDev]

@scaling01 Most teams don't need that level though. As long as the specs are clear, keep the tasks small, independently tested & verified, wiggum loop or similar with the best harness/model for each stage, and you are good. The model is now the least important bit but the most expensive.

[TomanatorDev @TomanatorDev]

@EddCoates It honestly sounds like you set it up insecurely. Flipped the switch on a WAF and assuming that it is then secure, without securing the underlying architecture. Kind of like putting your hand on your ass after you shit yourself.

[TomanatorDev @TomanatorDev]

@EddCoates Origin IP can be leaked through DNS, Email server or a misconfigured SSL cert. If not that, they already have malware in there. They simply aren't using advanced evasion as it costs too much. What value are they getting from you? I doubt it's scraping. More likely resources

[TomanatorDev @TomanatorDev]

@gsuchismit @Prathkum How did you manage the context window? Did you have to feed it the raw XLA HLO graph dumps or was the C++ enough for it to find the memory leaks?

[TomanatorDev @TomanatorDev]

@adhtri001 @RhysSullivan Not sure about benchmarks but I'm just going on experience. What other harnesses do you suggest for deepseek pro?

[TomanatorDev @TomanatorDev]

@Picco_Santi @JDWang11 @handlewest @PaulSolt What value is Claude giving you that others can't? They are all the same in my exp. Sure, some give better results with lazy input. When you need to get past QA, sonar, webinpect, pen test, perf test, then the cracks always show and lazy input doesn't work regardless of model

[TomanatorDev @TomanatorDev]

@quantumleap68 @JDWang11 @handlewest @PaulSolt No model is one shooting a production grade, performant and scalable app. The creators themselves don't even claim that

[TomanatorDev @TomanatorDev]

@Anaya_sharma876 Fedora just gets out of your way. It feels like it was built for developers. Everything is a shortcut or command away. It has podman and toolbox natively which are just beautiful. Toolbox is that good, developing without it feels like the wild west now. I can imagine going back

[TomanatorDev @TomanatorDev]

@plainionist Yes. No one wants to review or test. I've never had PRs with so much info but saying so little. I don't need how many lines or files were added/removed in the PR overview. But I do need to know why you put middleware inside middleware that has a recursive function with no exit 🤦

[TomanatorDev @TomanatorDev]

@mardehaym AI is just enabling laziness at scale. Spending a few hours on some hooks would stop this from ever happening. I can't understand how you could leave an agentic workflow running without guardrails in place. This outcome is inevitable. And it will inevitably happen again on Claude

[TomanatorDev @TomanatorDev]

@blader I'm curious what you're building that would warrant using it often? I can't see many actual good every day use cases that produce real value, especially considering the cost. I'd love to read up on good use cases where it is providing real value

[TomanatorDev @TomanatorDev]

@DaveShapi It is an LLM. It isn't intelligent. It predicts the next token based on the context. A single "stupid" or criticism outside the core task just warps the underlying geometry of the vector. It's Math. You are just reducing the probability of it staying on point. Then complaining 🤦

[TomanatorDev @TomanatorDev]

@DavidOndrej1 It's an LLM. Just a mathematical matrix. What benefit do you get out by inputting dumbass? I don't get that at all. It literally just predicts based on what you give and it's training. Just a waste of tokens, bandwidth and a quick way to poison the context for the cherry on top

[Wes Winder @weswinder]

fable 5 just refactored my entire codebase in one call

67 tool invocations. 1M+ new lines. 24 brand new files

it modularized everything. broke up monoliths. cleaned up spaghetti

none of it worked

but boy was it beautiful