HNHacker News
TopNewBestAskShowJobs

machiaweliczny

695 karma · joined July 12, 2015

submissionscomments
machiaweliczny··on Claude Shannon: Mathematician, engineer, genius and juggler (2017)
I’ve watched vide with guy that supposedly has 200IQ and in his view IQ is ability to focus on small on small set of things but many high IQ ppl can’t „refocus” it easily and aren’t doing well in life.
machiaweliczny··on Show HN: Hacker News frontpage as a print newspaper that you can personalize
How you compare a liked set to single embedding? By best match? Or is there some fancier way?
machiaweliczny··on Evaluating the world model implicit in a generative model
But if you squint then sensory actions and reactions are also sequential tokens. Even reactions can be encoded alongside input as action tokens and as single token stream. Anyone tried sth like this?
machiaweliczny··on Trump wins presidency for second time
There’s a reason that US has gun rights in constitution and it’s dems that want to take it away.

It’s interesting tradeoff that should in theory protect from tyranny.

It’s interesting what would happen in Russia if majority of population had guns.

machiaweliczny··on Trump wins presidency for second time
France seems to be doing better than UK it seems
machiaweliczny··on Trump wins presidency for second time
I see little analysis of Elon’s role in this election. Does anyone know if the accusation of him takin Russian money for Twitter were real?
machiaweliczny··on Show HN: Hacker News frontpage as a print newspaper that you can personalize
Nice. What algorithm is used for recommendation?
machiaweliczny··on Computer use, a new Claude 3.5 Sonnet, and Claude 3.5 Haiku
In normal critical system u use 3 CPUs. With LLM u can 1000 shot majority voting. Seems like approaches like entropix might reduce hallucinations also.
machiaweliczny··on Computer use, a new Claude 3.5 Sonnet, and Claude 3.5 Haiku
I'm waiting for Aider benchmark
machiaweliczny··on Computer use, a new Claude 3.5 Sonnet, and Claude 3.5 Haiku
Probably better than 4o-mini, 4o-mini isn’t great from my testing. loses focus after 100 lines of text
machiaweliczny··on Computer use, a new Claude 3.5 Sonnet, and Claude 3.5 Haiku
You can use Cursor (VS fork) with private Anthropic key
machiaweliczny··on Use Prolog to improve LLM's reasoning
Has anyone tried to mine LLM world model via sampling to extract all relations it's believes to be true (like 99%+ certain) into Prolog like clauses. I think this is way to achieve reliable world/domain models in logical sense (non-probabilistic). Probably brain doesn't do it but it could be cool anyway. Seems like good sampler could somehow mine this info by using stuff like I believe that X is true or false for all X imaginable. Then try go generate relations for these etc.
machiaweliczny··on Canvas is a new way to write and code with ChatGPT
Would love a more specific example of this. How to use this architect mode.

BTW surprised that tarballs work - aren’t these compressed?

machiaweliczny··on Canvas is a new way to write and code with ChatGPT
Just use aider - it’s CLI and open source. You can open terminal in IDE and you just use /add file for context then /drop to clean context. Better than Copilot and just less worse UX than Cursor (which has benefit of knowing which page you have open). There’s also Claude engineer but not tested it yet.

Forgot to mention killer feature - droping links to docs automatically fetches them in aider which helps with grounding for specific tasks

machiaweliczny··on If AI seems smarter, it's thanks to smarter human trainers
Yeah, I feel like you get much better answers when you pull enough context with leading context (but without insinuating answers).

I wonder if someone tried something like this: "Given user question generate prompt that tells LLM that he is a master in this subject and then ask the question" - something in this allows LLM to precompute what the focus of question will be so it can answer much better and it should be automated for one-off questions.

machiaweliczny··on Liquid Foundation Models: Our First Series of Generative AI Models
From my experience these large context are just techinical feasability but there needs to be a lot better internal benchamarks to claim it really works. I've tested on real world task and it all fails so far.
machiaweliczny··on Liquid Foundation Models: Our First Series of Generative AI Models
I've tested gpt4o and they've clearly improved since I've tested last year ago when poems were working only in english (like it would apply translation when you asked for other language)
machiaweliczny··on Learning to Reason with LLMs
In general if hallucination ratio is 2% can't it be reduced to 0.04% by running twice or sth like this. I think they should try establishing the facts from different angles and this probably would work fine to minimize hallucinations. But if this was that simple somebody would already do it...
machiaweliczny··on Ilya Sutskever's SSI Inc raises $1B
How about spiders intelligence? They don’t even have brain
machiaweliczny··on Ilya Sutskever's SSI Inc raises $1B
YeS, exactly. The trick is to have enough tough data so you find optimal one. I think as we will scale models back to smaller sizes we will discover viable/correct representations
machiaweliczny··on Ilya Sutskever's SSI Inc raises $1B
Me after watching channel5 I think some of it should go to poor people instead of billion dollars roulettes only. Thought the problem is with even richer corporations I feel and financial derivatives and not fully here.
machiaweliczny··on Ilya Sutskever's SSI Inc raises $1B
They might never work for novel discovery but that probably can be handled by outside loop or online (in-context) learning. The thing is that 100k or 1M context is a marketing scam for now.
machiaweliczny··on Ilya Sutskever's SSI Inc raises $1B
Happy to review this in 5 years
machiaweliczny··on Show HN: A discovery-focused search engine for Hacker News
Looks overwhelming on mobile
machiaweliczny··on Artificial intelligence is losing hype
I feel like currently the group feeling the most impact from Claude is frontend devs as tools like v0.dev, Claude etc. can almost create frontend from API schema (I've tested only page per page) which is great. It's probably because there's examples on the internet and that's why it works well.
machiaweliczny··on Fine-tuning now available for GPT-4o
I have sub to Claude now due to projects/artifacts (caching really) but aider with new gpt4o works well
machiaweliczny··on Fine-tuning now available for GPT-4o
I have 5 lines shell to load git into single file and using google AI studio. Tested it to simplify dreamcoder implementation but results are ~ so far
machiaweliczny··on Go is my hammer, and everything is a nail
Using esoteric lang constructs is bad idea IMO. Makes software harder to read/port/maintain. It’s useful for some hack & high perf stuff but not typical apps.
machiaweliczny··on Europe is in danger of regulating its tech market out of existence
Yes, it’s exactly what is needed in EU. Especially now as we can’t trust in trans-Atlantic relations. EU will go with own tech and military 100%
machiaweliczny··on Europe is in danger of regulating its tech market out of existence
The only difference is that US has lower cost of debt and that’s why they can buyout EU companies
← PreviousPage 5 of 29Next →