Juraj
@aksuno.bsky.social
Freelance software engineer, AI/ML researcher, and co-founder of https://vectorsight.net. Also rock-climber and occasional runner.
That’s the thing - I feel like Gemini works great without any specific prompt templates.
The key is to provide enough details so there’s little room for improvisation. Otherwise there’s no guarantee it will improvise the way you want it to.
The key is to provide enough details so there’s little room for improvisation. Otherwise there’s no guarantee it will improvise the way you want it to.
May 24, 2025 at 5:29 PM
That’s the thing - I feel like Gemini works great without any specific prompt templates.
The key is to provide enough details so there’s little room for improvisation. Otherwise there’s no guarantee it will improvise the way you want it to.
The key is to provide enough details so there’s little room for improvisation. Otherwise there’s no guarantee it will improvise the way you want it to.
Fair point. I had to check the release date as I feel it’s around for ages. Anyway, what I’m trying to say is that Google models work best for my use cases.
May 24, 2025 at 5:25 PM
Fair point. I had to check the release date as I feel it’s around for ages. Anyway, what I’m trying to say is that Google models work best for my use cases.
I second that. Still waiting for better coding model than Gemini 2.5 Pro. Ib terms of speed and code quality.
May 23, 2025 at 11:32 AM
I second that. Still waiting for better coding model than Gemini 2.5 Pro. Ib terms of speed and code quality.
What’s wrong with pydantic? 🤔
March 6, 2025 at 8:17 AM
What’s wrong with pydantic? 🤔
Interesting read.
Now I wonder what other capabilities can be improved using RL only. One could create bunch of different reward models and let them train models for longer time.
I know RL training is expensive, but even S1-like experiments could expose patterns and behaviors.
Now I wonder what other capabilities can be improved using RL only. One could create bunch of different reward models and let them train models for longer time.
I know RL training is expensive, but even S1-like experiments could expose patterns and behaviors.
February 9, 2025 at 11:41 PM
Interesting read.
Now I wonder what other capabilities can be improved using RL only. One could create bunch of different reward models and let them train models for longer time.
I know RL training is expensive, but even S1-like experiments could expose patterns and behaviors.
Now I wonder what other capabilities can be improved using RL only. One could create bunch of different reward models and let them train models for longer time.
I know RL training is expensive, but even S1-like experiments could expose patterns and behaviors.
Although there are few solutions that can do whole RAG on database side, you can definitely do hybrid search and reranking solely in SQL database.
SQLite (sqlite-vec) version
simonwillison.net/2024/Oct/4/h...
PostgreSQL (pgvector)
github.com/pgvector/pgv...
SQLite (sqlite-vec) version
simonwillison.net/2024/Oct/4/h...
PostgreSQL (pgvector)
github.com/pgvector/pgv...
Hybrid full-text search and vector search with SQLite
As part of Alex’s work on his [sqlite-vec](https://github.com/asg017/sqlite-vec) SQLite extension - adding fast vector lookups to SQLite - he’s been investigating hybrid search, where search results f...
simonwillison.net
January 17, 2025 at 7:03 PM
Although there are few solutions that can do whole RAG on database side, you can definitely do hybrid search and reranking solely in SQL database.
SQLite (sqlite-vec) version
simonwillison.net/2024/Oct/4/h...
PostgreSQL (pgvector)
github.com/pgvector/pgv...
SQLite (sqlite-vec) version
simonwillison.net/2024/Oct/4/h...
PostgreSQL (pgvector)
github.com/pgvector/pgv...
Oh wow! Gradio is going strong lately! 💪
January 16, 2025 at 11:03 PM
Oh wow! Gradio is going strong lately! 💪
Just tried it and searched for myself. The UI looks is really polished. And I like all the metrics it shown me.
Bravo! 👏
Bravo! 👏
January 16, 2025 at 5:28 PM
Just tried it and searched for myself. The UI looks is really polished. And I like all the metrics it shown me.
Bravo! 👏
Bravo! 👏
Didn’t know about that one. Looks great. Thanks!
January 16, 2025 at 5:23 PM
Didn’t know about that one. Looks great. Thanks!
Reposted by Juraj
LLMs are actually stateless. You put in the historic context as input.
Chat is only a UI constraint to build this context iteratively.
Chat is only a UI constraint to build this context iteratively.
January 4, 2025 at 8:53 PM
LLMs are actually stateless. You put in the historic context as input.
Chat is only a UI constraint to build this context iteratively.
Chat is only a UI constraint to build this context iteratively.
You just inspired me to write more! 🙂
January 3, 2025 at 6:25 PM
You just inspired me to write more! 🙂
I guess I could just read your blog, save my time and get summary everything. But it still interests me. - 2/2
January 3, 2025 at 6:19 PM
I guess I could just read your blog, save my time and get summary everything. But it still interests me. - 2/2
I have to ask as I was wondering about it for a while - how do you manage to produce so much content?
Like seriously. I would love to read a post about how you approach and research what’s new, what tool you made and use daily to consume all your feeds and sources. - 1/2
Like seriously. I would love to read a post about how you approach and research what’s new, what tool you made and use daily to consume all your feeds and sources. - 1/2
January 3, 2025 at 6:18 PM
I have to ask as I was wondering about it for a while - how do you manage to produce so much content?
Like seriously. I would love to read a post about how you approach and research what’s new, what tool you made and use daily to consume all your feeds and sources. - 1/2
Like seriously. I would love to read a post about how you approach and research what’s new, what tool you made and use daily to consume all your feeds and sources. - 1/2
No country for old men.
January 2, 2025 at 8:36 AM
No country for old men.
And then there’s also user prompt understanding. Average human is still a poor prompt engineer.
January 1, 2025 at 6:54 PM
And then there’s also user prompt understanding. Average human is still a poor prompt engineer.
It’s hard to answer, but as you pointef out - the data quality. That might include various distributions.
There is this paper that shows how dataset with 1k long instructions outperforms larger datasets of worse quality.
arxiv.org/pdf/2402.048...
There is this paper that shows how dataset with 1k long instructions outperforms larger datasets of worse quality.
arxiv.org/pdf/2402.048...
arxiv.org
January 1, 2025 at 6:53 PM
It’s hard to answer, but as you pointef out - the data quality. That might include various distributions.
There is this paper that shows how dataset with 1k long instructions outperforms larger datasets of worse quality.
arxiv.org/pdf/2402.048...
There is this paper that shows how dataset with 1k long instructions outperforms larger datasets of worse quality.
arxiv.org/pdf/2402.048...
Ended up with Contabo, hope will do the job.
December 31, 2024 at 4:30 PM
Ended up with Contabo, hope will do the job.