HNHacker News
TopNewBestAskShowJobs

jokethrowaway

3,052 karma · joined June 3, 2020

submissionscomments
jokethrowaway··on 30% drop in O1-preview accuracy when Putnam problems are slightly variated
I don't think llm generalise much, that's why they're not creative and can't solve novel problems. It's pattern matching with a huge amount of data.

Study on the topic: https://arxiv.org/html/2406.15992v1

This would explain o1 poor performance with problems with variations. o3 seems to be expensive brute forcing in latent space followed by verification which should yield better results - but I don't think we can call it generalisation.

I think we need to go back to the drawing board.

jokethrowaway··on Coconut by Meta AI – Better LLM Reasoning with Chain of Continuous Thought?
Definitely Twitter.

Some linkedin too.

jokethrowaway··on Coconut by Meta AI – Better LLM Reasoning with Chain of Continuous Thought?
This sounds like brute forcing a solution to make up for lack of intelligence.

In an IQ test, like the one in the arc agi test, a human sees the pattern instantly and effortlessly. o3 tries N paths until it stumbles on the right one and assess that there is a pattern.

I think we need a radically different architecture, this is a gimmick.

jokethrowaway··on Coconut by Meta AI – Better LLM Reasoning with Chain of Continuous Thought?
What can you learn from something parroting data we already have?

Similarly, we are now finding that training on synthetic data is not helpful.

What would have happened if we invested 1/100 of what we spent on LLM on the rule based approach?

jokethrowaway··on I Run LLMs Locally
I have a similar pc and I use text-generation-webui and mostly exllama quantized models.

I also deploy text-generation-webui for clients on k8s with gpu for similar reasons.

Last I checked, llamafile / ollama are not as optimised for gpu use.

For image generation I moved from automatic webui to comfyui a few months ago - they're different beasts, for some workflow automatic is easier to use but for most tasks you can create a better workflow with enough comfy extensions.

Facefusion warrants a mention for faceswapping

jokethrowaway··on Ask HN: Are you unable to find employment?
I know of many hiring manager who hire with recruiters that specialise in "diverse" candidates, where diverse means women and people of color only, so they can have better looking "diversity" stats. They certainly have a preference but this has been going since (at least) 2018 - it was true before the massive layoffs of 2023-24 and the end of ZIRP. What changed recently is that the market is really crap.

Incidentally all the people I know looking for a job are white males, but it may just be because they're the majority of people in tech I know anyway.

On legality of the above (given I was pretty sure it was illegal): It's legal, because "positive discrimination" is legal in the relevant EU countries (and in the UK).

jokethrowaway··on Why OpenAI's Structure Must Evolve to Advance Our Mission
We should abolish taxes, so real companies won't need to hide behind the non-profit status and real non-profits could shine.
jokethrowaway··on Fake Nintendo lawyer is scaring YouTubers, and its not clear YouTube can stop it
What about verifying the entity sending the request is legit?

An email @nintendo.com is not that hard to get for the legal team of nintendo

jokethrowaway··on Does current AI represent a dead end?
so maybe you want less mistakes?
jokethrowaway··on Does current AI represent a dead end?
copyrighting a sequence of numbers should have never existed in the first place

great if AI accelerates its destruction (even if it's through lobbying to our mafia-style protect-the-richest-company governments)

jokethrowaway··on Does current AI represent a dead end?
Nah, the trick with o3 solving IQ tests seems to be that they bruteforce solutions and then pick the best option. That's why calls that are trivial for humans end up costing a lot.

It still can't think and it won't think.

LANGUAGE models (keyword: language) is a language model, it should be paired with a reasoning engine to translate the inner thought of the machine into human language. It should not be the source of decisions because it sucks at doing so, even though the network can exhibit some intelligence.

We will never have AGI with just a language model. That said, most jobs people do are still at risk, even with chatgpt-3.5 (especially outside of knowledge work, where difficult decisions need to be taken). So we'll see the problems with AGI and the job market way earlier than AGI, as soon as we apply robotics and vision models + chatgpt 3.5 level intelligence. Goodbye baristas, goodbye people working in factories.

Let's start working on a reasoning engine so we can replace those pesky knowledge workers too.

jokethrowaway··on Are Immutable Linux Distros right for you?
Indeed, that's nice

Fixing the issue ends up being rather difficult though

jokethrowaway··on Are Immutable Linux Distros right for you?
This experience has been unique to (k)ubuntu (more than 15 years ago) for me.

I've been running rolling release distros for a decade and never had any problems - you have to follow some software migrations when needed, but I managed to migrate to systemd on Arch without an issue while any dist upgrade on ubuntu was wrecking my system.

jokethrowaway··on Are Immutable Linux Distros right for you?
I had the opposite experience because I want to run a lot of software in random repos.

I can make a nix-shell for each project but then every nix upgrade was forcing me to go through a lengthy reinstall + wrecking compatibility sometimes.

Not to mention the amount of derivations I had to write myself just to use latest packages.

Using things like virtualenv instead of nix-shell can fix the general instability, but packaging is too big of a problem.

I went back to Arch.

jokethrowaway··on Fighting spam with Haskell at Meta (2015)
Same! Well, Rust fits the bill too.

Generally, any language with a type system advanced enough.

jokethrowaway··on Fighting spam with Haskell at Meta (2015)
I don't see how that is language specific. I think it's entirely based on the project size.

Pick up docker or kubernetes and you'll have to learn plenty of stuff. Pick up ory/hydra to do auth in go, same thing.

Check out the age (from FiloSottile) or the rage (Rust port) codebase and they require similar level of understanding.

jokethrowaway··on Conda: A package management disaster?
my approach is to ignore all the *conda stuff and:

yay -S python-virtualenv # I'm on arch, do not confuse with 12 similarly named alternatives pyenv virtualenv 3.10 random-python-crap pyenv local 3.10.6/envs/random-python-crap pip install -r requirements.txt

and it works (sometimes deps are in some other places, or you have to pass -c constraints.txt or there is no file and you need to create it in various ways)

At least by not using local .env directories, I always know where to find them.

I install a lot of AI projecst so I have around 1TB just for the same python dependencies installed over and over again.

Sometimes I can get away with trying to use the same venv for two different projects but 8/10 deps get broken.

jokethrowaway··on Conda: A package management disaster?
easy if you start from scratch, hard if you want to get existing projects working

also it doesn't always work, I got stuck with some dependencies when it works it's amazing

jokethrowaway··on [dead]
Are they far right in the sense they are for economic freedom and against soviet style resource centralisation and redistribution - or far right in the sense they're racist and burning people in gas chambers?

We really need to start using the political compass.

jokethrowaway··on Veo 2: Our video generation model
Yes you can, with some limitations

https://github.com/Tencent/HunyuanVideo/issues/109

jokethrowaway··on Lfgss shutting down 16th March 2025 (day before Online Safety Act is enforced)
Indeed, or VATMOSS, which made all small merchants move their ecommerces to Amazon, Gumroad (and Paddle) to avoid the complexity
jokethrowaway··on Chatbot hinted a kid should kill his parents over screen time limits: lawsuit
I think no-one should be exposed to social media and doom scrolling.

Still it shouldn't be banned away, it's the parents responsibility to control access for their kids and everyone should be smart enough to avoid social media.

If you don't, that's on you, suit yourself.

jokethrowaway··on OnlyFans models are using AI impersonators to keep up with their DMs
Some do, some don't. Some people like their bosses, some don't. Some people like their clients, some don't.

It's not all black and white.

jokethrowaway··on Pgroll – Zero-downtime, reversible, schema changes for PostgreSQL (new website)
I don't think it's that bad but I used sqitch from the start of the project - so maybe I am the problem

I also don't think rebasing is the nightmare everybody makes it out to be. git rebase -i follow the instructions and you are good to go

jokethrowaway··on One of our clients hasn't paid us $130k – or "Why Every Contract Clause Matters"
I've been in this situation, a little bit of a legal threat does wonders

This is likely incompetence, not malice

jokethrowaway··on Dioxus 0.6 – Crossplatform apps with Rust
Or 10 european seniors
jokethrowaway··on Show HN: Cut the crap – remove the AI bullshit from websites
This is not a problem I feel, the nuggets of content are easy to find and read imho and I think there is a greater chance the AI will mess up the content's meaning. If you returned the full text of the article, that'd be cool.

Where your solution has potential is in removing the idiotic EU cookie banners, various useless popups, banners, obnoxious menus, autoplaying videos and what not.

If news websites were just a repository of text files, that would be great.

jokethrowaway··on Ask HN: Does retro computing make you feel younger or older?
It doesn't make me feel any kind of way.

Retro gaming is a necessity: modern games, like most modern software, are too good at manipulating the player's psychology and I certainly won't let my kids play free-to-play crap or let them on social media.

For myself: I genuinely like old games more than modern (with a few exceptions, eg. Balatro comes to mind)

I don't see a lot of appeal for retro programming; sure, dependency hell was not a problem in the past and tooling was simpler, but modern stacks have a series of productivity multipliers. You need to be careful in selecting the right technologies - but after that you are literally flying compared to old SDKs.

jokethrowaway··on Tokyo University Used "Tiananmen Square" Keyword to Block Chinese Admissions
A better example would be Harvard blocking people from Iran
jokethrowaway··on Wuhan lab samples hold no close relatives to virus behind Covid
Mhmm let's see what's more likely...

Zoonosis right next to a lab that does gain of function research on flu viruses - and not anywhere else in the world

OR

Lab leak covered by the CCP

← PreviousPage 6 of 34Next →