HNHacker News
TopNewBestAskShowJobs

sgt101

7,940 karma · joined January 24, 2012

nothing to see here, move along, move along.
submissionscomments
sgt101··on Complete hardware and software setup for running Deepseek-R1 locally. ($6000)
Hmm - if only macpro's had memory that could be upgraded like in the olden days.
sgt101··on Emerging reasoning with reinforcement learning
"My king, it's amazing these things can contain and regurgitate knowledge on demand. The next step is an army of Gollems that will destroy your enemies, all I need is three bags of gold." - Sargon the scribe, near Baghdad first month, 2024 BC

"My king, the Hittites have created scrolls that are wide and long despite that we burned their caravan taking palm leaves to the north. They have found a way to flense lambs to make pages themselves, and their ink is pure silver. They will soon have an army of Gollems and will destroy your empire. To stop this all I need is ten bags of gold." - Sargon the scribe, near Baghdad second month, 2024 BC

"My king, alas all the gold has gone, and yet neither will the gollems walk, nor have the Hittities come. My dancing girls grow cold, and I have no figs for my table all I need is 20 bags of gold" - Sargon the scribe, near Baghdad third month, 2024 BC

What happened to Sargon next is not recorded...

sgt101··on How has DeepSeek improved the Transformer architecture?
Compute has been more than doubling because people have been spending silly money on it. How long ago would a proposal for a $10m cluster for ML have been thought surreal by any funding agency? Certainly less than 10 years ago. Now people are talking of spending billions and billions.

Madness.

sgt101··on Nvidia, ASML Plunge as DeepSeek Triggers Tech Stock Selloff
hodl?
sgt101··on No Bitcoin ETFs at Vanguard (2024)
My wife can't wear a bitcoin. She can wear a gold necklace and will be pleased to do it even if the price goes to 10%
sgt101··on Emerging reasoning with reinforcement learning
I'm no neuroscientist, but I hung out with a few, so here are some bits I picked up from them. In biological brians:

- neurons have signal surpression (-ive activation) as well as propagation (+activation)

- neurons appear to process and propagation due to unknown internal mechanisms.

- there are complex sub structures within biological neural nets where the architecture is radically different from in other sections of the network, strongly in contrast to the homogenous structure of ANN's

- many different types of neurons with different properties and behaviors in terms of network formation and network activity are present in BNN's in contrast to ANN's

- BNN's learn during processing. ANN's are static after training.

sgt101··on Emerging Reasoning with Reinforcement Learning
We've had "reasoning" machines for a long time - I learned chess playing against a computer in the 1980's.

But we don't have reasoning that can be applied generally in the open world yet. Or at least I haven't seen it.

In terms of society it should be easy to track if this is true or not. Healthcare and elder care settings will be a very early canary of this because there is huge pressure for improvement and change in these. General reasoning machines will make a very significant, clear and early impact here. I have seen note taking apps for nurses - but not much else so far.

sgt101··on Emerging reasoning with reinforcement learning
Given the observations of the workings of brains though (slow, low energy) we're back to some magic to make this happen?
sgt101··on Emerging reasoning with reinforcement learning
It would be startling if humans could reason this way with two kilos of meat consuming 40w. Even more surprising given that we get answers in ~1s with a cross brain comms time of about 0.25s

To me it's clear that human reasoning is different from a massive search of a latent space. I can say that when I am thinking I maybe try half a dozen ideas or scenarios, but very rarely more. I can't say where those ideas come from or how I make them up though. Maybe we can't frame what it is and how it works with human languages though, which might make it seem magical in some way.

Or maybe there's a good framing that I don't know - would love to learn!

sgt101··on Coping with dumb LLMs using classic ML
Options:

Finetune the models to be better

Optimise the prompts to be better

Train better models

sgt101··on FIPA: Good old fashioned agent standards
One thing we did with Zeus which remained proprietary was to build a graphplan planner to replace the STRIPS planner. This was better, but unfortunately it only shifted right by one step - very quickly we found that there were planning issues in the agent loop that were intractable.

I think that SAT planners could help further, and I think that LLM's can play a role by suggesting subplans as well. We know that for closed domains like games RL based planners like Alphago can be spectacular, but for the kind of open domain reasoning that Agents have to do I don't think that there is a solution yet.

Would be cockahoop if that's not true !

sgt101··on FIPA: Good old fashioned agent standards
Interesting, I do wonder about FIPA performative acts. On the one hand they have a clear function in terms of tagging steps in the protocols, and the protocols have a very clear function in terms of co-ordination and resource management. On the other hand this doesn't seem to have been the driver for the development of the performative act library - rather they developed them from the perspective of "let's classify the intent of all messages".

I did a paper for AAMAS 20 years ago where I reviewed the use of performative acts by real developers. No one in my project would co-author it with me (bad sign) and it got rejected (doh) but I do believe that it was useful and it's a shame that it didn't get into the literature. Anyway, the punch line was that when devs actually tried to use FIPA they just used the performative acts in the protocols that they implemented, and those protocols were almost always Fipa-Request and Fipa-Query.

I've created some mirrors as well.

FIPA mirror : https://github.com/sgt101/FIPA-Mirror-/tree/main

and

Zeus: https://github.com/sgt101/zeus

Zeus has a FIPA transport as well as its native transport, it's got reasoning engines that it uses to act as approximately a BDI agent. The problem is that its STRIPS style planner can't reason so much... and also the RETE engine has sharp limits around rule chaining explosions and so on. The other big problem is that I haven't actually run it successfully for about 10 years due to Java versioning challenges (it works on 1.4 very well... however...)

sgt101··on Stargate Project: SoftBank, OpenAI, Oracle, MGX to build data centers
What is the evidence for 1) ? I thought that the latest models were getting "somewhere" with fairly trivial reasoning tests like ARC-1
sgt101··on The quiet rebellion of a little life
Ok, no capital letters as a writing affectation is immediately upsetting... but I'll let my anxiety about that subside and focus on the content.

I get it, I really do. I find it a daily struggle. Last week I told my boss I couldn't travel this week. He was fine about it, there wasn't even a discussion. But I've worried about it ever since. Not because I want a raise or a promotion, but because I need this job and I worry that I won't get another good one if I lose it. I need it to cover my families needs, I need it so that I can live a good life post work. I need to think about post work because we all know that when your hair goes white and you struggle with the accessibility of the office you're not going to get another one.

So, there is no compact that will create the feeling of safety and fulfillment that this article wishes for. Just money.

sgt101··on UK's hardware talent is being wasted
Yup - I'm not saying it's good or fair, or bad and unfair. I'm just saying this is what it is.
sgt101··on UK's hardware talent is being wasted
>I think most Doctors etc need to wait until they're consultants until they make decent money.

A consultant gets £100k -> £140K ish from the NHS. However, many supplement that with private work and therefore make significantly more.

sgt101··on Norepinephrine-mediated slow vasomotion drives glymphatic clearance during sleep
might be a very long, very deep sleep...
sgt101··on Aaron Swartz and Sam Altman
Proof? :) where has proof ever been found in truth?

It's not proof, but there is some good evidence out there. Here are two interesting papers I found informative.

https://arxiv.org/html/2403.04121v2

https://arxiv.org/html/2411.12580v1

sgt101··on Aaron Swartz and Sam Altman
>Second, if the crime was the act of scraping then it’s directly comparable. But if the crime is publishing the data for free, that’s quite different from training AI to learn from the data while not being able to reproduce the exact content.

They often do reproduce the exact context; in fact it's quite a probable output

>“Probabilistic plagiarism” is not what’s happening or even aligned with the definition of plagiarism (which matters if we’re talking about legal consequences). What’s happening is that it’s learning patterns from the content that it can apply to future tasks.

That's what I think people wish would happen. Sometime they have been shown to learn procedural knowledge from the training data but mostly it's approximate retrieval.

sgt101··on Ask HN: Books about people who did hard things
I came here to say this. Props.
sgt101··on Predictions Scorecard, 2025 January 01
I feel a counter is that hyping and going along with hype leads to substantial misallocation of capital and this leads to human misery.

How much money has been burned on robo-taxis which could have been spent on incubators for kids.

sgt101··on Agents Are Not Enough
Hi - have a look at this book if you are interested [1] (Mike Wooldridge, Multi-Agent Systems)

[1] https://amzn.eu/d/6a1KgnL

Here are Mike's credentials :https://www.cs.ox.ac.uk/people/michael.wooldridge/

sgt101··on Agents Are Not Enough
Hi - have a look at this book if you are interested [1] (Mike Wooldridge, Multi-Agent Systems)

[1] https://amzn.eu/d/6a1KgnL

Here are Mike's credentials :https://www.cs.ox.ac.uk/people/michael.wooldridge/

sgt101··on Agents Are Not Enough
Hi - have a look at this book if you are interested [1] (Mike Wooldridge, Multi-Agent Systems)

[1] https://amzn.eu/d/6a1KgnL

Here are Mike's credentials :https://www.cs.ox.ac.uk/people/michael.wooldridge/

sgt101··on Agents Are Not Enough
Hi - have a look at this book if you are interested [1] (Mike Wooldridge, Multi-Agent Systems)

[1] https://amzn.eu/d/6a1KgnL

Here are Mike's credentials :https://www.cs.ox.ac.uk/people/michael.wooldridge/

sgt101··on AI and Startup Moats
Nearly complete security isn't security. If the potential is there people will find it, other models will find it.

Everythings fine until one day $200m disappears from your balance sheet and no one can explain why!

sgt101··on AI and Startup Moats
(total side track) There are other things that some customers want though:

- for the recommendations to offer me things I want or need, not things I just bought

- to be able to evaluate the quality of items rather than just the price of items

- for Amazon to extend it's brand around the items that I buy. "Amazon Recommends" is just so weak and offers no assurance or opportunity for loyalty. It's more or less meaningless and I suspect it's something that suppliers buy.

As every in business it's very difficult. I know that Amazon is humongous and knows it's business inside out. I am sure that Amazon insiders just feel tired reading other people's ideas about what would make things better, but on the other hand I do think that the narratives of business inevitability (and AI inevitability) are just false. Yes they have triumphed until recently, but what's happening in China really does undermine the idea that the future will be everyone just grifting to everyone else for a dime while the big corps enshitify anything that emerges from the primordial ooze.

Not that I think that what's happening in China is good.

sgt101··on A messy experiment that changed how I think about AI code analysis
I completely disagree.

Computer Science has a massive ethics crisis. Uncritical adulation and a total lack of accountability or consequence is part of that.

There is a massive misallocation of capital which is burning opportunity for our society. Users are getting terrible experiences and systems because people read this sort of thing and believe it. Trust in our technology is eroded and this has consequences for actual people. We have abandoned the standards that protected people and you have the view that these standards, or a shadow of them even, are unnecessary?

Someone has taught a generation that this is all ok, it isn't.

sgt101··on A messy experiment that changed how I think about AI code analysis
Yeah, it is hard.

The principal you need to work to is that you need to create evidence that other people will find compelling and then show that you have interrogated your results to show that you have checked that it's really working better than chance and not the result of some fluke or other. Finally you need to find a way to explain what's happening - like an actual mechanism.

1. Find or make a data set - I've been using code_search_net to try and study the ability of LLM's to document code, specifically the impact optimising n-shot learning on them*, this may not be close enough to your application, but you need many examples to draw conclusions. It's likely that you will have to do some statistics to demonstrate the effects of your innovations so you probably need around 100 examples.

2. Results from one model may not be informative enough, it might be useful/necessary to compare several different models to see if the effect you are finding is consistent or whether some special feature of a model is what is required. For example, does this effect work with only the largest and most sophisticated modern models, or is this something that can be seen to a greater or lesser effect with a variety of models?

3) You need to ablate - what is it in the setup that is most impactful? What happens if we change a word or add a word to the prompt? Does this work on long code snippits? Does it work on code with many functions? Does it work on code from particular languages?

4) You need a quantitative measure of performance. I am a liberal sort , but I will not be convinced by an assertion that "it worked better than before" or "this review is like an senior, I think". There needs to be a number that someone like me can't argue with - or at least, can argue with but can't dismiss.

*I couldn't make it work, I think because the search space for finding good prompting shots (sample function)is vast and the output space (possible documents) is vast. Many bothans died in order to bring you this very very very (in hindsight with about $200 of OpenAI spending) obvious result. Having said that I am not confident that it couldn't be made to work at this point so I haven't written it up and won't make any sort of definitive claim. Mainly I wonder if there is a heuristic that I could use to choose examples a-priori instead of trying them at random. I did try shorter examples and I did try more typical (size) examples. The other issue is that I am using sentence similarity as a measure of quality, but that isn't something I am confident of.

sgt101··on Court strikes down US net neutrality rules
Cue FAANG buying telcos.

Effectively, AT&T returns (by a very strange and expensive route).

It will be very interesting to see what effect this has on networks globally.

← PreviousPage 12 of 34Next →