Javi™

Right now, I'm building AI products with a small team that outperforms what used to require an entire department. Agents handle code, research, testing, and deployments. We handle vision, architecture, and the hard decisions. This is the most exciting time to be a software engineer!

Local models are definitely a step in the right direction. Cloud models have been necessary while hardware evolved, but they also come with latency, cost and privacy issues. I think there’s another shift coming. As architectures stabilize, it starts to make sense to move from general-purpose compute GPUs, to more specialized silicon, with models increasingly baked closer to the hardware (as companies like Taalas are starting to show). I’m convinced that AI will eventually come embedded in our iPhone and other devices, becoming immediate, private, and virtually free to use.

Local models are the way. @AlexFinn is right. "Normies" will just get them automatically. Consumers will care about privacy and reliability, two things you can only get from local models. The "Pros" are moving to open models (I saw it at NVIDIA GTC where nerds waited in line to buy a $5,000 DGX inferencing box to get it loaded up with local models. Why? The cloud models (GPT, Gemini, Grok, Claude) are a little bit better but cost the pros TONS every day. I met more than one developer at GTC that's spending more than $1,000 a day in tokens to build their software. This shift is underway even though most don't recognize it yet.

I'm curious about the linting process. Do you just use LLMs to verify the sources, or you have more sophisticated mechanisms? I'm finding that after some time working with a knowledge base, it starts accumulating errors that propagate to all derived resources. With a large number of originals, it becomes challenging to know or verify what's true and what's not. This structure is proving really useful, but I think this is one of the main challenges we need to solve to achieve longer unsupervised research.

LLM Knowledge Bases Something I'm finding very useful recently: using LLMs to build personal knowledge bases for various topics of research interest. In this way, a large fraction of my recent token throughput is going less into manipulating code, and more into manipulating knowledge (stored as markdown and images). The latest LLMs are quite good at it. So: Data ingest: I index source documents (articles, papers, repos, datasets, images, etc.) into a raw/ directory, then I use an LLM to incrementally "compile" a wiki, which is just a collection of .md files in a directory structure. The wiki includes summaries of all the data in raw/, backlinks, and then it categorizes data into concepts, writes articles for them, and links them all. To convert web articles into .md files I like to use the Obsidian Web Clipper extension, and then I also use a hotkey to download all the related images to local so that my LLM can easily reference them. IDE: I use Obsidian as the IDE "frontend" where I can view the raw data, the the compiled wiki, and the derived visualizations. Important to note that the LLM writes and maintains all of the data of the wiki, I rarely touch it directly. I've played with a few Obsidian plugins to render and view data in other ways (e.g. Marp for slides). Q&A: Where things get interesting is that once your wiki is big enough (e.g. mine on some recent research is ~100 articles and ~400K words), you can ask your LLM agent all kinds of complex questions against the wiki, and it will go off, research the answers, etc. I thought I had to reach for fancy RAG, but the LLM has been pretty good about auto-maintaining index files and brief summaries of all the documents and it reads all the important related data fairly easily at this ~small scale. Output: Instead of getting answers in text/terminal, I like to have it render markdown files for me, or slide shows (Marp format), or matplotlib images, all of which I then view again in Obsidian. You can imagine many other visual output formats depending on the query. Often, I end up "filing" the outputs back into the wiki to enhance it for further queries. So my own explorations and queries always "add up" in the knowledge base. Linting: I've run some LLM "health checks" over the wiki to e.g. find inconsistent data, impute missing data (with web searchers), find interesting connections for new article candidates, etc., to incrementally clean up the wiki and enhance its overall data integrity. The LLMs are quite good at suggesting further questions to ask and look into. Extra tools: I find myself developing additional tools to process the data, e.g. I vibe coded a small and naive search engine over the wiki, which I both use directly (in a web ui), but more often I want to hand it off to an LLM via CLI as a tool for larger queries. Further explorations: As the repo grows, the natural desire is to also think about synthetic data generation + finetuning to have your LLM "know" the data in its weights instead of just context windows. TLDR: raw data from a given number of sources is collected, then compiled by an LLM into a .md wiki, then operated on by various CLIs by the LLM to do Q&A and to incrementally enhance the wiki, and all of it viewable in Obsidian. You rarely ever write or edit the wiki manually, it's the domain of the LLM. I think there is room here for an incredible new product instead of a hacky collection of scripts.

@BettaTech Cuando nadie se acuerde de cómo funciona la web, quizás la propia AI cree sus propios frameworks y discuta con otras AIs sobre que framework es mejor por nosotros…

El sector del desarrollo siempre ha tenido sus problemas, pero ahora mismo parece un parque infantil. Añoro las discusiones de si React es mejor que Angular, y nunca pensé que diría esto jaja

@davidmarcus @openclaw The biggest limiting factor I see to achieve that is the accumulative errors on the agent’s memory when they run unsupervised for a while.

We’re likely < 12 months from unsupervised software development. Not just better models. Full closed loops: generate → run → evaluate → fix → repeat. Using @openclaw you can already see it. Once loops + models improve together, supervision will stop making sense.

@Shpigford @BearNotesApp

What app has the best markdown reading/writing experience you’ve ever experienced in your entire life?

@jeffrey_way Tests can probably prevent it because they’re a deterministic steering system, but you probably need to understand what you’re doing to create the right ones… and at that point I doubt you can still call it “vibe coding” 🤔

I think it's probably undeniable that RIGHT NOW, if you effectively vibe code your projects, the codebase will become increasingly worse with each passing day. Not sure that any combo of tests + formatters + skills can prevent that.

@rand_longevity I doubt it. The work is infinite. The 9-5, 5-days-a-week schedule comes from what the average human is capable of sustaining. It’s not like there’s exactly 40 hours of work per week. If we could work 48 or 500 hours a day and still have a life, we’d do it... Well, now we can!

your full time 9-5 job is gonna go down to 3 days a week soon

@DorotheaBaur Indeed, that’s the whole point of AI! A tool to process unreliable inputs and produce useful-ish outputs. We have code for anything that needs to be exact.

It bears repeating: „LLMs are inherently probabilistic, not deterministic. They try to guess the answer that sounds best, not the objectively correct one. And that’s not a problem to be fixed. It’s fundamental to how these models work.“

@emollick More than a bottleneck, it’s a ceiling, and that’s probably a good thing, it’s the sign we’re still part of the game 😅

A sign that human creativity is a bottleneck is that this year everyone can generate almost any image or video they can think of for nearly free and the April Fools posts are basically just as bad as any other year.

@awesomekling Maybe to build 100+ projects in parallel?

I wonder if anyone will assemble a team of 100+ paid software engineers ever again.

@juanmacias Habrá que ver exactamente qué es lo que pasa, pero desde luego está claro que el código ya no es un valor diferenciador. Justo estaba leyendo esto y creo que ha acertado en muchas cosas: x.com/michaelxbloch/…

A todos los que pensáis que la IA no va acabar con el Saas… estáis muuuuy equivocados. Solo aquellos con integraciones, acceso exclusivo a APIs, red de partners, van a sobrevivir… el resto? Lo dudo y no por el tema precio

@pmitu You probably don’t want an AI model processing your wire transfers 😅

Why would you build anything non-AI-powered these days?

@pmddomingos Because LLMs are just approximate text processors. Everything else is needed to put together the text they need to process and maximize the accuracy for specific use cases.

If LLMs are so smart, why do they need all these prompts, harnesses, post-training, scaffolding, etc.?

We've been trying something that partially aligns with this model, and the main limiting factor in our case is the reliability of our “world model” because when you stack agents, the hallucinations compound over time. Everyone needs to be aware that some things may be “made up” and must verify them. That verification need is the new bottleneck. That said, a non-reliable world model is already showing a ton of value because it surfaces info that you didn’t even notice before.

@wickedguro I tried to run some automated social media experiments, but the posts tend to be a bit soulless, and people generally ignore or reject them because it’s obvious they were generated by AI. Have you noticed that kind of issue with Postiz?

@DotCSV Supongo que por esto el cli de Claude Code tiene tantos cientos de miles de líneas de código 😅

Usad /buddy en Claude Code y mostradme vuestra mascota legendaria! Si a alguno le sale un pangolín que me lo pase con el cable link 👍

@konnydev Deciding what to do 😜

If everything is done by AI in the future, what is left for us?

@sachintwtss We're using Vercel and Railway a lot lately.

AWS is too overwhelming. Heroku is too dead. need a platform that's just right. What are you deploying on?

@NaivaidyaY66600 aifindr.ai

Founders, what are you working on this week? Drop your product's link here 👇👇

@boxmining It increases the value of AI operators’ labor. So the key to survival is to embrace the technology.

AI doesn’t reduce the value of money. It reduces the value of labor. Big difference.