• 0 Posts
  • 66 Comments
Joined 1 year ago
cake
Cake day: June 7th, 2023

help-circle
rss
  • I think your job in your current form is likely in danger.

    SOTA Foundation Models like GPT4 and Gemini Ultra can write code, execute, and debug with special chain of thought prompting techniques, and large acale process verification on synthetic data and RL search for correct outputs will make this 10x better. The silver lining to this is that I expect this to require an absolute shit ton of compute to constantly generate LLM output hundreds of times for each internal prompt over multiple prompts, requiring immense compute and possibly taking longer than an ordinary software engineer to run. I suspect early full stack developer LLMs will mainly be used to do a few very tedious coding tasks and SWEs will be cheaper for a fair length of time.

    I expect it will be 2-3 years before this happens, so for that short period I expect workers to be “super-productive” by using LLMs in the coding process, but I expect the crossover point when the LLM becomes better is quite soon, perhaps in the next 5 years as compute requirements go down.


  • I suppose having worked with LLMs a whole bunch over the past year I have a better sense of what I meant by “automate high level tasks”.

    I’m talking about an assistant where, let’s say you need to edit a podcast video to add graphics and cut out dead space or mistakes that you corrected in the recording. You could tell the assistant to do that and it would open the video in Adobe Premiere pro, do the necessary tasks, then ask you to review it to check if it made mistakes.

    Or if you had an issue with a particular device, e.g. your display, the assistant would research the issue and perform the necessary steps to troubleshoot and fix the issue.

    These are currently hypothetical scenarios, but current GPT4 can already perform some of these tasks, and specifically training it to be a desktop assistant and to do more agentic tasks will make this a reality in a few years.

    It’s additionally already useful for reading and editing long documents and will only get better on this end. You can already use an LLM to query your documents and give you summaries or use them as instructions/research to aid in performing a task.


  • Current LLMs are manifestly different from Cortana (🤢) because they are actually somewhat intelligent. Microsoft’s copilot can do web search and perform basic tasks on the computer, and because of their exclusive contract with OpenAI they’re gonna have access to more advanced versions of GPT which will be able to do more high level control and automation on the desktop. It will 100% be useful for users to have this available, and I expect even Linux desktops will eventually add local LLM support (once consumer compute and the tech matures). It is not just glorified auto complete, it is actually fairly correlated with outputs of real human language cognition.

    The main issue for me is that they get all the data you input and mine it for better models without your explicit consent. This isn’t an area where open source can catch up without significant capital in favor of it, so we have to hope Meta, Mistral and government funded projects give us what we need to have a competitor.



  • NFTs are stupid AF for most of the tasks people currently use them for and definitely shouldn’t be used as proof of ownership of physical assets.

    However, I think NFTs make a lot of sense as proof of ownership of purely digital assets, especially those which are scarce.

    For example, there are several projects for domain name resolution based on NFT ownership (e.g you look up crypto.eth, your browser checks that the site is signed by the owner of the crypto.eth NFT, then you are connected to the site), as it could replace our current system, which has literally 7 guys that hold a private key that is the backbone of the DNS system and a bunch of registrars you have to go through to get a domain. This won’t happen anytime soon but it is an interesting concept.

    Then I think an NFT would also be good as a decentralized alternative to something like Google sign in, where you sign up for something with the NFT and sign in by proving your ownership of it.

    In general though I find NFTs to be a precarious concept. I mean the experience I’ve had with crypto is you literally have a seed phrase for your wallet, and if it gets stolen all your funds are drained. And then for an NFT, if you click on the wrong smart contract, all your monkeys could be gone in an instant. There is in general no legal recourse to reverse crypto transactions, and I think that is frankly the biggest issue with the technology as it stands today.


  • “I use Signal to hide my data from the US government and big tech”

    “Wait, you seriously still use Reddit? Everyone switched to the Fediverse!”

    “Wow, can’t believe you use Apple! Android is so much better.”

    No one who isn’t terminally online understands what these statements mean. If you want people to use something else, don’t make it about privacy and choose something with fancy buttons and cool features that looks close enough to what they have. They do not care about privacy and are literally of the mindset “if I have nothing to hide I have nothing to fear”. They sleep well at night.






  • This is another reminder that the anomalous magnetic moment of the muon was recalculated by two different groups using higher precision lattice QCD techniques and wasn’t found to be significantly different from the Brookhaven/Fermilab “discrepancy”. More work needs to be done to check for errors in the original and newer calculations, but it seems quite likely to me that this will ultimately confirm the standard model exactly as we know it and not provide any new insight or the existence of another force particle.

    My hunch is that unknown particles like dark matter rely on a relatively simple extension of the standard model (e.g. supersymmetry, axioms, etc.) and the new physics out there that combines gravity and QM is something completely different from what we are currently working on and can’t be observed with current colliders or any other experiments on Earth.

    So probably we will continue finding nothing interesting for quite some time until we can get a large ML model crunching every single possible model to check for fit on the data, and hopefully derive some better insight from there.

    Though I’m not an expert and I’m talking out of my ass so take this all with a grain of salt.






  • I think this is downplaying what LLMs do. Yeah, they are not the best at doing things in general, but the fact that they were able to learn the structure and semantic context of language is quite impressive, even if it doesn’t know what the words converted into tokens actually mean. I suspect that we will be able to use LLMs as one part of a full digital “brain”, with some model similar to our own prefrontal cortex calling the LLM (and other things like vision model, sound model, etc.) and using its output to reason about a certain task and take an action. That’s where I think the hype will be validated, is when you put all these parts we’ve been working on together and Frankenstein a new and actually intelligent system.



  • On the config point, I understand NixOS has got the trade off with the config specification, but I was mainly talking about the need for a GUI manager for the config so that regular people could use it. Like it automatically has a flake setup that works with multiple hosts, is synced automatically, has a store app where you can click add on a package then rebuild, etc. Mainly, I don’t want to have to specify gnome on or KDE off, I want to click a button that does that, and for all other things you have to put in the flake. It could also have a little popup window for more advanced config that requires actually editing the code.


  • For the love of God please stop posting the same story about AI model collapse. This paper has been out since May, been discussed multiple times, and the scenario it presents is highly unrealistic.

    Training on the whole internet is known to produce shit model output, requiring humans to produce their own high quality datasets to feed to these models to yield high quality results. That is why we have techniques like fine-tuning, LoRAs and RLHF as well as countless datasets to feed to models.

    Yes, if a model for some reason was trained on the internet for several iterations, it would collapse and produce garbage. But the current frontier approach for datasets is for LLMs (e.g. GPT4) to produce high quality datasets and for new LLMs to train on that. This has been shown to work with Phi-1 (really good at writing Python code, trained on high quality textbook level content and GPT3.5) and Orca/OpenOrca (GPT-3.5 level model trained on millions of examples from GPT4 and GPT-3.5). Additionally, GPT4 has itself likely been trained on synthetic data and future iterations will train on more and more.

    Notably, by selecting a narrow range of outputs, instead of the whole range, we are able to avoid model collapse and in fact produce even better outputs.