HN user
Research done by Prof. Subbarao on reasoning & planning with LLM seems to point in the negative.
https://www.youtube.com/watch?v=uTXXYi75QCU
He has lots of good threads distilling his research. https://twitter.com/rao2z
Ways of Being: Animals, Plants, Machines: The Search for a Planetary Intelligence
by James Bridle. James pulls an brilliant move of using various examples in technology as a tool to help us rethink our place in nature & reconnect with the (beyond human) world.
He weaves a fascinating tale with myriads of threads such as Cybernetics, Neural Nets, Internet, Random Numbers, Analog computers, Slime molds, Sortition vs Voting, Mycelium, Mysticism in animals, Turing machines, Personhood of non-humans etc.
A nice summary to whet your appetite, https://pluralistic.net/2022/06/07/more-than-human/#umwelt
Search/Reasoning/Inference time compute, however you phrase it is still essential. You need search to improve upon learning to work in novel situations.
Can't unseen the amount of double negatives in this article.
I am working on RL and robotics. I came across Levin in Lex's podcast. And then went on a binge of his other podcast appearences. I agree totally with you, I would very much like to build agents that adapt to different circumstances like "simple organisms". I am not familiar with biology, but I plan to build competence here to follow Levin's work to a point that I could potentially collabrate with biologists or learn from their work. Any suggestions (books etc) that would be salient towards this goal is much appreciated!
Have you checked RSSM approach in DreamerV1,V2,V3,PlaNet? It uses deterministic (GRU hidden state) and discrete stochastic latent states. The deterministic and stochastic (sampled) latent state are used to predict the next state. I think the stochastic state might help with your problem a bit.
One interesting research group tackling this question is ex-computer engineer, currently biologist Prof. Micahel Levine's group.
They built xenobots (biological robots) from frog cells which are not evolutionarily programmed to do tasks which they accomplished. Their research in part deals with emergence of agency at multi-cellular level. And they want to get at the fundamental signatures of what makes some collection of agents into a new single agent at an higher level.
Checkout, https://youtu.be/XheAMrS8Q1c
And a new research paper (inhavent had the time to read yet) on this question of how coherent higher order agents emerge from underlying sub-agents.
https://www.frontiersin.org/articles/10.3389/fpsyg.2019.0268...
Worth checking out farcaster (farcaster.xyz). They are working towards a solution where things are "sufficiently decentralised". The server admins can't prevent users from following one another etc.
https://www.varunsrinivasan.com/2022/01/11/sufficient-decent...
Farcaster(farcaster.xyz) is working towards sufficiently decentralised social network (https://www.varunsrinivasan.com/2022/01/11/sufficient-decent...). It aims to have a model where server admins don't get the power to prevent users from following one another.
Does apple even encrypt the actual image data on device? Their system document says "payload = visual derivative + neural hash" and only that is encrypted with a secondary level encryption. And they didn't go through with e2ee for icloud last I heard. This elaborate system makes no sense if they very well could have done it in cloud.
It feels like elaborate privacy theatre trojan horse to introduce in device surveillance.
His narrative pushing feels nefarious to me. I think he is using his new tech podcast (all-in) as a new platform to build a image of a pragmatic do-gooder billionare (to whomever who would buy that shtick), spin a new narrative and test out few ideas like (remove capital gains tax, he running for governorship) for approval.
If GME debacle taught us (again?) anything, it's that markets are moved by narratives. And those raccoons who control the narrative can move the market towards their favour.
Great post on epsilon theory on this https://www.epsilontheory.com/hunger-games/
I have experience in deep learning research/engineering in NLP and am getting into ml for robotics.
Really excited for this course + hardware. I think one top value for me personally is that more time could be spent on hacking the ml algorithms, replicating and comparing with real research work (https://www.duckietown.org/archives/63665).
I live in India and find the shipping costs of MOOC Founder's Edition Kit (369 USD + 49 USD) a bit expensive. Just the Duckiebot MOOC Founder's Edition costs 14 USD to ship.idk what custom duty would the govt put on it :(.
1. Is there any alternative to shipping it? 2. Is there anyway to build similar bot from parts for the course? 3. I guess, since you're non-profit it's made such that it's cost effective rather than for profit. It would be great if you describe why you chose each of the component for the current bot.
If uber/lyft doesn't allow their drivers to price themselves and not treat them as employers, they are more or less a cartel rather than a marketplace connecting buyers and sellers of a service.
Maybe they can show a optimal price and allow drivers to change that and charge drivers for the customers they get from the platform.
Same thing happened to me. I was using org-mode for GTD and note-taking for the past six months. I shifted all my notes to org-roam with links between them. But then I tried roam, it's UX really won over me.
One thing was accessing my notes with all roam features in all the devices. I don't have to setup dropbox/syncthing/whatever in every machine I own and install emacs with my settings every time. And mainly mobile, If I use orgzly, I don't get any benefit of org-roam in it. So I gave up and embraced roam cloud.
I think there is a huge gap to be filled for a device agnostic emacs that works well on mobile (touch), browser, native and is programmable.
Masks can protect other people primarily. If you're sick and don't know (asymtomatic), developing symptoms only now, if you're sick and don't have any option but to go out etc.
And your point on masks not protecting incoming droplets is wrong. When people sneeze, droplets can go more than 5 feet. Masks reduce exposure to such droplets.
https://medium.com/@Cancerwarrior/covid-19-why-we-should-all...
Research is on going on how immunity responds to decreased exposure to viral loads.
And masks can remind people to not touch faces and remind others that we're in a pandemic.
Wouldn't Land Value Tax solve the housing problem?
Would you consider using neural nets to solve integration and differential equations as it doing system 2 reasoning?
https://ai.facebook.com/blog/using-neural-networks-to-solve-...
Or what tasks are in the domain of system 2?
Lex, you're doing an amazing job. The level of commitment and hard work you put in to make this podcast is incredible.
On the technical depth side criticism you're getting, I would say the one with Ian Goodfellow had optimal balance.
And don't take destructive criticism about bad attitude or tone etc by some commentors . For someone introverted it's already draining to put yourself in front of a camera ...
But why is microsoft experimenting with the toy language then?
https://www.infoq.com/news/2019/11/microsoft-exploring-rust-...
70% of microsoft errors would be gone in future if they start using what you demean as "toy" language.
https://www.zdnet.com/article/microsoft-70-percent-of-all-se...
This one isn't a Cartel as its like a taxi company but co-owned by workers. But Uber could be considered a cartel because they don't consider the car drivers as workers, so they are effectively setting prices for the market of drivers.
I think allowing extensions on existing classes/structs can also solve this problem without having structural typing.
ie the client can create a new interface and explicitly implement it for existing classes/structs.
I think Swift and rust allows this.
Rust.
Feynman and Einstein comes to mind.
This post by Micheal Nielsen explores his strategies to convert stuff into Anki in detail.
http://augmentingcognition.com/ltm.html
Also this on math into Anki.
It's ironic that these people who consider or signal(if you're cynical) virtue through their jobs completely ignore the larger picture.
Facebook is a cesspool which wants to lock people in the web into a single monopoly. And worse it is to prey on the users attention by dealing cognitive equivalent of drug hits.
Where were these woke engineers when Zuck tried to push his drug into developing nations in ruse of "free internet", or inumerable other breach of user trust and privacy dor sake of more profit?
Its inexcusable that when you're working in a fundamentally immoral/ammoral organization, you try to signal virtue hard. Its like a foot soldier in a mob teaching ethics lessons when cleaning blood off his shoes. Better to be honest about where you get your pay than be "holier than thou".
How far have we gone from the light of visionaries like Alan Kay and others who paved foundation to the web and computing, who wanted technology to magnify the intellect ...
Allennlp is one such project for NLP. Uses dependency injection to separate out the components driving for maximum re-usability. https://github.com/allenai/allennlp
What if we had regulations to force service providers to provide APIs and interoperability in their services? What if messaging someone on Facebook from Google hangouts was as easy as sending an email?
Exactly, the kind of regulation that would promote real competition than silos of shitware built on network effects.
API Enforcement, Net neutrality, Transparency, Free software enforcement, Copy left and privacy regulation are pro-consumer, pro-user, pro-individual and anti-authoritarian. These are the regulations that would promote freedom instead of standing in the way or worse grab the notion of freedom completely.
Stallman has been preaching about it for ages. It is sad that people who fight for regulation don't have faith on the individual, or want to provide the tools for betterment of the individual (a centrally controlled internet with official version of the truth is never the answer).
One point I do lean in on the pro-regulatory side of things. The platforms showing automated feeds/suggestions have to share the same culpability as individual doing it. i.e. If a goverment deems an individual can't do X, or has to pay Y in damages, then a platform promoting the video actively (hosting it without a searchable link is fine) through automated feeds is culpable of sharing it.
I am not promoting govt regulation, but just saying companies should not be treated as exceptions.
Or use emacs.
I started using Org mode in emacs. I use it to record the next tasks (Todos), expected results, and actual results. Its damn convinent that I can run code inside org file and say "cat" out the result of an experiment output as JSON by another script into my notebook. I can even embed and run multiple languages using org-babel.
I think I will be coming up with workflows and checkboxes soon for organizing a experiment and easily integrating results of the experiments I am running from disparate software.
I think its very easy to incorporate the stuff in this post, and create a new template in org mode for this. And its so flexible that it will fit all usecases.
Emacs is too awesome.
On a similar vein, "Kenneth Stanley: Why Greatness Cannot Be Planned: The Myth of the Objective".
https://www.youtube.com/watch?v=dXQPL9GooyI&app=desktop
Kenneth leads the AI research in Uber. He had worked on Evolutionary strategies for machine learning, artificial Life etc.