HNHacker News
TopNewBestAskShowJobs

piterrro

525 karma · joined November 20, 2012

https://logdy.dev Real-time logs viewer https://leadjobs.dev/ Engineering Leadership job board

posinsk () gmail.com

submissionscomments
piterrro··on P99 0 ms* autocomplete for 240M domain names
I just typed a random sequence of the characters, long enough to be certain such domain doesnt exist. No only, the browser send an autocomplete request for every keystroke but for each request it returned a set of proposed domain names (which I'm 100% certain doesnt exist). At this point, how do I understand which results are legit and which are fake? Also, it would be nice to highlight the typed part in the result set so I can visually see what matches exactly.
piterrro··on SQLite as a Document Database (2020)
I do that in psql and it works really well. But your post got me thinking since I need to find a solution to store content of multiple documents, have a way to do FTS as well as vector similarity. I dont need that for all of the documents at once - I need to do it either for one document or at most couple of documents.

Now I'm thinking I could have a separate database file per "batch", store it in object storage and then download on demand and query it as I want. This way I'll not bloat my primary storage size as well as I dont need a special vector DB since sqlite vector search will be enough for up to 50k vectors.

piterrro··on Show HN: A lightweight, stateless database for agent memory
I get the closed source open binary approach - I would test it if I woild be in a need!
piterrro··on RAG Is Simpler Than You Think
RAG only makes sense if you have an LLM review the results, pick the most relevant ones and iterate further if there's a need running another query and repeating the process. Raw dump of vector search (even with reranking) is asking for troubles (or rather weird user questions like 'why this crap popped up in the results?')
piterrro··on Show HN: Huzzah – a novel approach to coding with AI
You could write that pseudocode as a prompt for the agent and get the same result. Use plan mode to understand what agent wants to do.

Am i missing anything?

piterrro··on PostgreSQL for Everything
true to that - currently using psql (in a single monolithic codebase) as: sql db, json db, vector store, logs store, full-text search, queue, message bus.

multiple processes connected to it.

piterrro··on Geolocating a random island using geometry and CUDA programming
really impressive, could that be the way to locate yourself without GPS? assuming we know more/less where we are
piterrro··on Cloudflare's AI Psychosis
Companies specialized in building DCs and AI labs specialized in niche models?
piterrro··on Cloudflare's AI Psychosis
The problem is, beside their size, they have not moat to compete in AI space. Their data centers are spread across the world in other companys DCs (btw do they own any DC actually from top to bottom?).
piterrro··on Don't classify, hallucinate
I would propose the following, query vector store for 10 closest categories based on a query, feed it to an LLM, in the prompt ask it to produce a single digit 0-9 representing the number of the most appropriate choice. Use plain text prompt, dont inflate token count with JSON. There you go, you just drastically reduced the output pricing.

Additionally you could experiment with a reranker instead of an LLM or after reranking take top-3 results and then feed to LLM as input in order to reduce input token costs.

piterrro··on Choose Boring Technology (2015)
https://grugbrain.dev/ Similar on this topic
piterrro··on Mistral OCR 4.1
You can even run it on your desk if you want, a single gtx 4090 is enough. It can be fully air gapped.
piterrro··on Mistral OCR 4.1
Accuracy can have different dimensions, depends on what you can tolerate and whether you can detect it to apply more powerful methods.

Imagine you have a cheap and 99% accurate ocr. The other 1% you can detect and apply more powerful (more accurate but slower and more expensive) ocr method. What would you use? At scale these things add up.

piterrro··on Mistral OCR 4.1
Most use cases dont need that kind of accuracy, just doesnt justify the 3-4usd range. I build for that exact case (tender documents, we’re processing north of 100k pages per day), it doesnt need to recognize scanned written text from 1930s, its usually pdf/docs/scanned printed pages. The accuracy is great, bounding boxes are must have for proper grounding for building answers by LLMs. Tesseract was too slow and not enough in some cases (for example tables or images which we also recognize and describe)
piterrro··on Mistral OCR 4.1
For anyone interested, I have an ocr pipeline running on rented GPUs, doing around 1000pages for 0.05-01 usd with around 0.8 seconds per page with full bounding boxes support for grounding.

If you’re interested you can find contact to me via this profile.

3.5 usd/1000 pages is just too expensive…

piterrro··on Don't be a meat proxy
In these situations I always use AI to respond
piterrro··on Why we write our own C and C++ inference engines
Could this vllm port be faster to install? Im starting gpu machine multiple times a day and it takes 5 minutes to set vllm up. If Inise this port that time is minimized?
piterrro··on CRM: An open-source, agentic-first CRM
Just cant believe AI slop is getting 1k stars on github in two days. Nothing really innovative about this project, nothing you cannot do yourself in two days with a coding agent.

I wrote it before, OSS is changing in a way that will basically make it useless.

piterrro··on Linux on ESP32
When can we run doom on it?
piterrro··on Running Kimi K3 on a M1 Max
Will it fit on ESP32??
piterrro··on You Could Have Come Up with Kimi Delta Attention
At first I felt bad about not having come up with this solution. But then I realized I have problems with writing binary search by myself in JS and immediately felt better.

Now way I could have come up with Kimi Delta Attention.

piterrro··on Scriptc by Vercel: TypeScript-to-Native compiler, no JavaScript engine in binary
this is enough to keep them on the first page of HN and similar services. Its a growth strategy: invest tokens, build some "nice" project nobody wants, get some reach through publishment. Rinse and repeat.

Heads up, in 12 months from now, 90% of open source projects will be vibecoded, without actual users - only looking "interesting" (if not already?).

Its easy to vibe code a full blown compiler these days, the question is: will it be maintained? A catchy title is not enough to get sticky users, long term maintenance, community around the project is (or at least used to be).

Having said that, the only viable reason for these kind of projects I see is getting attention for a moment, boosting reach of Vercel logo across the web and then disappearing (or stop development).

If Vercel would think seriously about this project: they would show skin in the game - for example implement it as their own experimental runtime.

piterrro··on Scriptc by Vercel: TypeScript-to-Native compiler, no JavaScript engine in binary
That's a one big hell of a contribution ;-) https://github.com/vercel-labs/scriptc/commit/725b931cf43ba7...
piterrro··on It's getting harder to focus every day
This feels so familiar. Since more than a year or so Im witnessing more and more content being produced. One year ago or so it was enough to visit HN main page once a day and be up to day. These day, it needs a visit every few hours, the main page changes constantly. Not only that but also the amount of content received via newsletters.

Worst is with LLMs and coding agents, I give it a task to work on then go browsing the web, the LLM finishes in 2 minutes but I come back after 10…

At the end of the day, my mind feels super tired even though its the LLM that did the heavy lifting. I also barely remember what I was browsing 15 minutes ago. At this point I can honestly say - Im in trouble.

piterrro··on Show HN: AI Law Tracker – one audited API for US, EU and global AI law
Recently whenever I see something that's vibecoded and I see nobody taking personal responsibility for it: company entity, linkedin profile links - I leave immediately - it's not worth spending my time if the author was not willing to put their reputation at stake.

I guess this is a universal method. In the time when we can have 100 pages vibecoded in an hour by agents, how do we decide which are actually worth human time?

The agent doesn't feel a notion of cost or time, for them producing slop is free. For us humans, reading slop is at cost of cognitive overload.

piterrro··on Precursor
I think in 10/20 years from now, access to the Internet will be allowed only upon personal identification. Every website will be allowed to ask about your identity upon serving any content. Thats the only way I see this is going. The internet as it is right now does not have a future if majority of traffic will be done by agents. Thus, the cost of that traffic will have to be put on the users and since displaying ads doesnt make sense to agents, a paid access will be introduced (which is what cloudflare is slowly doing)
piterrro··on Show HN: Microsoft releases Flint, a visualization language for AI agents
For a while now, everytime I see „for AI agents” Im looking for benchmarks with comparable solutions.

The only „for AI agent” solution that will gain adoption is/should be measured by two dimensions: token usage and correctness.

If the solution doesnt use less tokens than generating chart.js code for example - why should I use it?

Same for correctness, if the generated chart spec is correct only 90% of the time - why should I use it?

It’s still early but I think this is the direction we should be thinking about „for AI agents” libraries and projects.

piterrro··on Building relationships with customers through support didn't turn out as hoped
From my experience building rapport only makes sense with your most valuable users in terms of revene. Everything else is noise. Non-paying users are most demanding. Its great to talk to lots of people in mvp or ppmf stage, after that once you nail your icp and start charging, you should make the human support a paid feature.

I’ll tell you even more, in enterprise b2b saas, usually the company paying few thousand per month would have less questions and requests than the one paying few hundreds.

piterrro··on Designing DB partitions you don't have to babysit
Oh, I meant bits not bytes. You are right, the uuid in psql is 16 bytes, the bigint is 8 bytes.
piterrro··on Mouse: Precision Editing Tools for AI Coding Agents
So I’ll have to buy a license for using my mouse now?
← PreviousPage 2 of 8Next →