HNHacker News
TopNewBestAskShowJobs

razodactyl

563 karma · joined September 8, 2014

submissionscomments
razodactyl··on [Discussion] OpenAI to pause use of "Scarlett Johansson" voice but
Seems like a great way to further poke the ants nest of media hype! Any publicity is free marketing.

Isn't the bigger deal the fact that we literally have "Her" as a technology and such a debate is now possible to begin with?

I've learnt that it's smart to take a step back with all the hype fueled by the media and realise all the competition means we get to see the technology as a whole (in form of open source etc.) down the line.

razodactyl··on GPT-4o
In my view, this was in response to the machine being colourblind haha
razodactyl··on GPT-4o
I think this is a great example of the bootstrapping that was enabled when they pipelined the previous models together.

We do this all the time in ML. You can generate a very powerful dataset using these means and further iterate with the end model.

What this tells me now is that the runway to GPT5 will be laid out with this new architecture.

It was a bit cold in Australia today. Did you Americans stop pumping out GPU heat temporarily with the new model release? Heh

razodactyl··on Spring Update
Reckon we can get in some last minute speculation? You know, in the spirit of last minute crunch developer demos.
razodactyl··on Spring Update
DAMMIT. You beat me by 2 minutes haha
razodactyl··on Google Quantum AI
Heh. This is an interesting comment. Imagine if we didn't know about relativity - we would have discovered it as an annoyance/weird quirk instead as we ran into it.

Reminds me of that story about the self-evolving chip that was tasked to learn how to classify tones and instead took advantage of specific flaws in its own package.

razodactyl··on SEQUOIA: Exact Llama2-70B on an RTX4090 with half-second per-token latency
Meta seem to be thinking 10 years ahead where anyone can run these models at the edge.

Perhaps it's not about where the model is hosted but what can be built on it.

Meta have added Llama3 across the board on all their apps.

Chat is fun but in the wrong context it's useless. However the training data return on millions of users is something interesting to pay attention to.... Llama4 might be a significant jump!

Increased model intelligence and innovative applications of language technology will be where the real value appears. Open-sourcing and allowing public amplification of abilities and enhancements is a very smart move.

The marketing department is also commendable. What happened to Grok? LLMs are everywhere - we're running them on home computers, that's where we should be pondering the next moves.

razodactyl··on SEQUOIA: Exact Llama2-70B on an RTX4090 with half-second per-token latency
There are always going to be pros and cons. That's why solutions like managed databases are reality. From an expert perspective it seems like there's more to lose but from the perspective of a company with employee turn over, possible data loss, security etc. the benefits start to far outweigh the costs.

This reasoning can mostly be applied here. If you want to learn about and pull the LLM apart. Perhaps fine-tune and tinker then 100% go ahead running locally. You however won't be able to scale this up easily for a consumer base and the electricity use and heat output starts to become a problem.

At some point it's more beneficial to pay the provider for inference, this includes upkeep, latest models, faster generation, stability, hosting etc.

Pros and cons! Choice is important and Meta is doing the right thing by the AI community and tech community in general by being realistic with these programs. The ecosystem is giving back by being able to access these high quality models.

razodactyl··on Let's Think Dot by Dot: Hidden Computation in Transformer Language Models
I think you may need to read my comment again. It's even mentioned in the summary:

> In summary, our results show that additional tokens can provide computational benefits independent of token choice. The fact that intermediate tokens can act as filler tokens raises concerns about large language models engaging in unauditable, hidden computations that are increasingly detached from the observed chain-of-thought tokens.

The paper discusses an unexplained benefit of additional computation regardless of which token is selected be it symbols or Lorem Ipsum.

I didn't mention anything about training. I'm speaking based on how the Transformer architecture itself is designed.

The "unauditable" computation is simply a result of how machine learning models work. The extra computation made available is mentioned in my explanation.

Keen to hear your thoughts on it though.

razodactyl··on GPT-4.5 or GPT-5 being tested on LMSYS?
https://imgur.com/a/z39k8xz

Seems like it's not too hard an ask.

razodactyl··on Let's Think Dot by Dot: Hidden Computation in Transformer Language Models
I get the feeling this works due to the following:

1. An input is processed with answer generated token-by-token.

2. The model can output based on probability the answer or a filler token. Low probability answers are ignored in favour of higher probability filler tokens (I don't have a better token to answer with than .....)

3. At a certain point, an alignment is made with what was learnt previously triggering a higher probability of outputting a better token.

This intuition being that I've noticed models respond differently based on where in context information appears: Can't speak for different embedding methods however as I'm sure this changes my thoughts on above.

If instead chain of thought prompting is used, the tokens further generated may interfere with the output probability.

So further to this, I'm thinking filler tokens allow for a purer ability for a model to surface the best answer it has been trained on without introducing more noise. Or we can use methods that resample multiple times to find the highest outputs.

These LLMs are practically search engines in disguise.

razodactyl··on Keep Out – WebGL Game
It's broken on iOS. It asks to turn to landscape but the screen doesn't expand.
razodactyl··on Lawyer divorces wrong couple through dropdown selection error
Yet we have to type the name of our repo in GitHub to confirm a deletion...

Imagine the headaches if a life event happened in between the divorce and un-divorce. Oops.

razodactyl··on [dead]
10:27pm
razodactyl··on Rerank 3: A new foundation model for efficient enterprise search and retrieval
Imagine you have 100 documents in a database and you "query" the documents and return 20 candidate results.

Similarity gave you 20 results but Re-ranking sorted those results further providing relevance.

That 4K is per document.

Edit: With sorted relevance, you can drop the lower scoring documents according to the model's confidence that the information in the subset is adequate to answer the query.

razodactyl··on Groq CEO: 'We No Longer Sell Hardware'
Absolutely on the mark with this comment. LLMs aren't magical end-goal technology. We have a while to go it seems before they've settled into all the use-cases and we've established what does and doesn't work.
razodactyl··on Ask HN: How to secure website for public launch
I feel I'm not supposed to upvote this as much as I have...
razodactyl··on The jobs being replaced by AI – an analysis of 5M freelancing jobs
And then you permute some variables and you're able to sweep through the space of every voice that ever existed.... Hyperbole for the point I'm trying to make - We've essentially trained the Million Monkeys already.
razodactyl··on Documentation for the AMD 7900XTX
It's not just this though. AMD cards are priced better and NVIDIA is overpriced due to the gap.

I would love to use AMD for my ML experiments and I would love to see healthy competition benefit this field.

razodactyl··on Ask HN: What is the current (Apr. 2024) gold standard of running an LLM locally?
I miss when redditors were really nice xD That was a while ago now. (Or maybe people just make me more nervous these days lol)
razodactyl··on Rhea, AI
However your feedback is still useful. Noted!
razodactyl··on Rhea, AI
Sorry, I just fixed that bug. Should be available now.
razodactyl··on Rhea, AI
Hi all,

I've been working on a project and thought I should start sharing.

Looking for a bit of feedback and suggestions from the brilliant minds of HN.

razodactyl··on Someone has been attempting to DDoS us for weeks and we do nothing
I like this a lot: Why? Because the attacks are directed to someone who isn't bothered and wastes their own resources. I've been a Table Plus user for near a decade now and enjoy the simple but highly compatible software they provide.
razodactyl··on Amazon spends another $2.7B on Anthropic
Theoretical end goal which will ensure a company's future but reality is never straight forward.

If AGI is achievable (I'm thinking likely given the right circumstances) then we'll see the same end result just like the current open source models floating around everywhere.

That said. I highly recommend checking out what Cohere's up to. Their Command-R model is pretty good and their infra. is fast.

razodactyl··on Emad Mostaque resigned as CEO of Stability AI
I can shed some light on this phenomena: These models are trained on many images but no thought is put into the "generalisation" aspect the ML community was so obsessed with during the deep-learning era.

Whatever Kinbaku is... haha

I think a key part that's missing currently is the agent training approach: https://youtu.be/v3UBlEJDXR0?si=8w4Jt0bNEBfIXkZl

It's very easy to create a "Stochastic Parrot" but I'm quite sure these models are capable of learning underlying information such as correct layout of a knot - given the right data and curriculum of course. Maybe slight architecture tweaks.

I'm sure this is the reason we're starting to see a normal amount of fingers or ability to write text. Proof of concept was 2015 until 2022 now we're starting to see interesting things come out of the workshops.

razodactyl··on Emad Mostaque resigned as CEO of Stability AI
I think it became apparent when mixtral came out. I've noticed too during training that my model overwrites useful information so it makes sense for these types of models to have emerged.
razodactyl··on Emad Mostaque resigned as CEO of Stability AI
Yep. It feels like a 20B parameter model.
razodactyl··on Emad Mostaque resigned as CEO of Stability AI
I approve this comment.
razodactyl··on Emad Mostaque resigned as CEO of Stability AI
This comment distracted me as I tried to recall a product I saw my mother using when I was younger. I fell asleep after I found it and never got back here until just now: Microsoft Works

This one was so obscure to find because it seems to exist in a weird space of being Microsoft Office but not.

← PreviousPage 8 of 13Next →