visarga

19 ms·

391.
by visarga 9mo ago
A quick glance over the 200 LOC impl - I see no error handling. This is the core of the agent loop, you need to pass some errors back to the LLM to adapt, while other errors should be handled by the code. There is also no model specific cod
392.
by visarga 9mo ago
Do you think the SDD approach is fundamentally wrong, or that Amazon's implementation was at fault?
393.
by visarga 9mo ago
No, I think we can get by with using CommonCrawl, pulling every few months the fresh content and updating the search stubs. The idea is you don't change the entry points often, you open them up when you need to get the fresh content. I
394.
by visarga 9mo ago
I had an idea - take SerpAPI and save top-10 or 20 links for many queries (millions), and put that in a RAG database. Then it can power a local LLM do web search without ever touching Google. The index would just point a local crawler towar
395.
by visarga 10mo ago
I am personally working on formalizing the design stage as well, the core concepts being Architecture, Goal, Solution and Implementation. That would make something like the complexity of an algorithm an explicit decision in a graph. It woul
396.
by visarga 10mo ago
> Anyone who understands how easy it is to copy bits should know that the original intent of copyright can't work anymore. AI makes this even more stringent. You cannot protect the "vibe" of your works, AI can replicate i
397.
by visarga 10mo ago
> But if I didn't already have experience of code review, I'd be limited to vibe-coding (by the original definition, not even checking). Code review done visually is "just vibe testing" in my book. It is not something
398.
by visarga 10mo ago
> This is what code review is for. My point is that visual inspection of code is just "vibe testing", and you can't reproduce it. Even you yourself, 6 months later, can't fully repeat the vibe check "LGTM" s
399.
by visarga 10mo ago
Why train to pedal fast when we already got motorcycles? You are preparing for yesterday's needs. There will never be a time when we need to solve this manually like it's 2019. Even in 2019 we would probably have used Google, solv
400.
by visarga 10mo ago
Everyone is so fixated on the output as the commodity, whether it’s a blog post or a piece of code, that they fail to see the interaction itself as the locus of value. You can still do your rewarding work in a chat session, it can force you
401.
by visarga 10mo ago
Not all AI generated outputs are slop, usually it's the low effort prompts that create slop. When you bring in external data or extensive human curation it is almost certainly not slop. I think many people put all AI outputs in the slo
402.
by visarga 10mo ago
I'd be happy if the original expression remains protected but abstractions are open to reuse. Right now a work blocks more than its specific expression, a whole space is forbidden around it, like fan fiction. In other words what is nee
403.
by visarga 10mo ago
In the present moment, time passes slow when something boring or painful happens, and passes fast for exciting or pleasant experiences. But in hindsight, time passes slow in periods where you had many new experiences and is almost missing i
404.
by visarga 10mo ago
I think the missing ingredient is not something the LLMs lack, but something we as developers don't do - we need to constrain, channel, and guide agents by creating reactive test environments around them. Not vibes, but hard tests, the
405.
by visarga 10mo ago
> I don't think we've seen any groundbreaking new research or architecture since the introduction of inference time compute ("thinking") in late 2024/early 2025 circa GPT-o4 It was model improvements, followed by
406.
by visarga 10mo ago
A generic wrapper is not a moat, but the context is. Both the LLM provider and the wrapper provider depend on local context for task activities. The value flows to the context, the LLMs and wrappers are commodities. Who sets the prompts sta
407.
by visarga 10mo ago
> The real question isn't "can OpenAI justify its valuation" but rather "what's the blended tax rate of its investor base?" Was that an organic "it's not A, it's B" or synthetic?
408.
by visarga 10mo ago
I like your concept of indexing high quality sources for RAG. For many queries we might not need the usual search engines.
409.
by visarga 10mo ago
> which is crazy, given how many resources it's sucking Gentleman, the dog writes poetry and music, but it is boring, mediocre quality. Overhyped dog.
410.
by visarga 10mo ago
So, what is the solution? Senior engineer looks over PR and signs LGTM? That is just "vibe testing". The worst kind of testing. I think the author is right, setting up tests to form a reactive environment for coding agents will le
411.
by visarga 10mo ago
> The dynamic would be totally different if LLMs actually brodged the brain-computer barrier and enabled near-frictionless generation of programs that match an arbitrary specification. Software engineering would change dramatically, but
412.
by visarga 10mo ago
Haha, I can't read long articles anymore because I want to reply, a habit I picked chatting LLMs.
413.
by visarga 10mo ago
The prefix tuning approach was largely abandoned for LoRA, it does not change the process if you tune the prefix or some adapter layers, but it is more flexible to train the LoRAs. The Skills concept emerged naturally when you see how codin
414.
by visarga 10mo ago
If we did that it would be much more expensive, keeping all weights in SRAM is done by Groq for example.
415.
by visarga 10mo ago
At the end of the day, you either validate every line of code manually, or you have the agent write tests. Automate your review.
416.
by visarga 10mo ago
Every error is a signal you need better tests. You can let the LLM create tests for every error it stumbles into, besides all the regular tests it can write on its own. Add all test scenarios you can think of, since you are not implementing
417.
by visarga 10mo ago
I would just put a PR_REVIEW.md file in the repo an have a CI agent run it on the diff/repo and decide pass or reject. In this file there are rules the code must be evaluated against. It could be project level policy, you just put your
418.
by visarga 10mo ago
You should take into consideration the time it took to make those 9200 tests originally. If you have good test coverage the agent can go much farther ahead.
419.
by visarga 10mo ago
Of course there is - if you write good tests, they compress your validation work, and stand in for your experience. Write tests with AI, but validate their quality and coverage yourself. I think the whole discussion about coding agent relia
420.
by visarga 10mo ago
You 100% need to test work done by AI, if it's code it needs to pass extensive tests, if it's just a question answered, it needs to be the common conclusion of multiple independent agents. You can trust a single AI as much as a H

More ›