Joseph Szymborski @JSzym

A great joint work with my brilliant student @JSzym. See the thread below and read our new paper in @NatMachIntell on data leakage in pLMs, when used for protein-protein interaction. ⬇️ 🔗 rdcu.be/e3PGD #mcgill #llm @Mila_Quebec #bioinformatics

Emad's COMBINE Lab @ McGill University @COMBINE_McGill

5 months ago

🧵 1/ 🎉New paper alert! Pretrained protein language models (#pLMs) are all the hype, but are they really helping us predict protein-protein interactions? 🤔Dive into our thread to see why you should read the full study @NatMachIntell. ⬇️ 🔗  rdcu.be/e3PGD

1 0 1 77 0

Do fish sleep? Yes! (they also yawn) - Our work on the genetics of widespread temporal niche partitioning in cichlids: biorxiv.org/content/10.110…

Emad's COMBINE Lab @ McGill University @COMBINE_McGill

5 months ago

🧵 1/ 🎉New paper alert! Pretrained protein language models (#pLMs) are all the hype, but are they really helping us predict protein-protein interactions? 🤔Dive into our thread to see why you should read the full study @NatMachIntell. ⬇️ 🔗  rdcu.be/e3PGD

1 0 1 77 0

@MorleyTheodore @mcgillu This is adorable.

This is one of the best descriptions I've lately ran into.

INTREPPPID - An Orthologue-Informed Quintuplet Network for Cross-Species Prediction of Protein-Protein Interaction biorxiv.org/cgi/content/sh… #biorxiv_bioinfo

@JennyENicholson do you have opinions on barrels of snakes? This place had both barrels and racks.

keyboard_arrow_left Previous keyboard_arrow_right Next

The future of #AI governance is here. Help us shape AI for the benefit of people, not just corporations. Join a global community of policymakers, engineers, and more in ensuring that AI remains open and accessible to all ➡️ open.mozilla.org/letter

Imagine if, in 1999, instead of DoJ bringing an antitrust suit against Microsoft over Internet Explorer, DoC started looking into whether it should be illegal to make open indexes of the web.

Soumith Chintala @soumithchintala

3 years ago

In 270 days, the Department of Commerce will determine whether they will allow open-weights or not. if you support open model weights and want something actionable to do, then figure out how to lobby your opinion to them.

33 152 670 469K 127

I'm excited to show a preprint of a paper my colleagues and I have been working on for quite some time! Link: medrxiv.org/content/10.110… We aimed to evaluate potential causes of variable performance integrating PRS into clinical prediction -- more details below! 👇

Whilst @geoffreyhinton is a *brilliant* scientist, scientists are *not* the people to teach us about risk management. The project manager for the Nuclear Information Project has already warned us about these "muddled analogies", leading to a "calorie-free media panic".

Geoffrey Hinton @geoffreyhinton

3 years ago

@ylecun Let's open source nuclear weapons too to make them safer. The good guys (us) will always have bigger ones than the bad guys (them) so it should all be OK.

180 79 839 905K 84

I have a family member studying virology. Everyone in the course is learning this forbidden knowledge that LLMs are not allowed to pass on. So why do we allow the teaching of virology, but do not allow an understanding of virology to be more accessible?

Nikki Teran @DrNikkiTeran

3 years ago

Will releasing the weights of large language models grant widespread access to pandemic agents? Turns out, yes, probably. 1/5

56 91 426 632K 328

The reason that (2) is in question and open source AI R&D is threatened is (1). Those questions are not disconnected. If you ask "give me arguments for why turbojets and rocket engines won't kill people", I could respond by pointing to a dozen treatises on how to design and build reliable turbojets and rocket engines. Those treatises don't exist for superhuman AI because superhuman AI has not been invented yet (and people who think scaling up Auto-Regressive LLMs will get use there are just wrong). So the question becomes what will superhuman AI systems look like? What is the blueprint? My answer to this (which you've heard many times) is that future AI systems, just like intelligent animals, will be *objective driven*. Whatever output or actions they produce will *have* to satisfy a set of objectives, some of which will be guardrails that guarantee their safety and controlability. We will set those objectives and guardrails, and AI systems will not be able to violate them *by construction*. These systems will stay under out control because their primary objective will be to be subservient to us. Now, we are not going to build super-intelligent objective-driven AI systems right away. We are going to build very simple systems first, and make sure the guardrails we put in place guarantee their safety. Then, we'll work our way up, building smarter and smarter systems, adjusting the objectives and guardrails as we go. It will be like turbojet and rocket engine reliability: an iterative process of refinement and careful engineering. Much of the x-risk hyperventilating is due to two fallacies: (1) just because a system is intelligent it will want to take over, (2) the minute we turn on a slightly flawed super-intelligent system, it will take over and destroy humanity. Number (1) is simply false. The desire to dominate is not linked to intelligence. That's not even true within the human species (the people who seem most motivated to dominate are not the smartest among us). To want to dominate, you need a hard-wired drive to do so. Only a few social species have such a thing (including humans, baboons, chimps, wolves, etc). Non social species (like orangutans) don't. But if you are still scared, we can hard-wire objectives into AI systems to make them submissive to humans (within limits). Number (2) reveals a complete misunderstanding of how new technologies are developed and deployed. The same way we don't build a giant turbojet airliner and put 1000 passengers in it on the first trial, we won't build a super-intelligent AI prototype and let it lose on the first try. We'll build cat-level AI first and figure out what guardrails are necessary as we work our way up the intelligence level.

Rule: Any theory about risk of harm, no matter how fringe and as long as it evokes irrational fear in the public, will be coopted by politicians, promoted, then used to seize more power.

European Commission @EU_Commission

3 years ago

Mitigating the risk of extinction from AI should be a global priority. And Europe should lead the way, building a new global AI framework built on three pillars: guardrails, governance and guiding innovation ↓

367 422 1K 2.9M 375

Fact check: extinction risk from AI, much like extinction risk from self-replicating grey goo nanobots or from an alien invasion from Kepler-62, should be pretty far down the list of global priorities. Maybe look at the less fictitious risks of climate change, pandemics, and war.

European Commission @EU_Commission

3 years ago

Mitigating the risk of extinction from AI should be a global priority. And Europe should lead the way, building a new global AI framework built on three pillars: guardrails, governance and guiding innovation ↓

367 422 1K 2.9M 375

I've spent the last few months interviewing >60 experts in law, economics, AI, alignment, etc, on the impacts of AI, and safety interventions. Today I'm publishing my first article, showing regulation designed to increase AI safety may backfire badly! fast.ai/posts/2023-11-…

Perfect illustration of why ML/AI is an empirical field, built on intuition. Next time you get lost in a sea of math while reading a paper, remember that the author probably discovered the method this way and tacked on the math later (to pass peer review).

Prafull Sharma @prafull7

3 years ago

History of attention discussed by @karpathy at ICVSS.

6 66 496 182K 167

@NateSilver538 @robinhanson I think there is an important distinction to make between "experts on Deep Learning" and "experts on AI ethics". While the opinons of the former are worth hearing, I think we fall into a trap of credentialism when we defer to high profile DL experts when AI safety experts exist.

@Lang__Leon @NateSilver538 @timnitGebru @DAIRInstitute She has a lot to say about existential risk, even though she argues it's not a primary concern.

This is a wonderfully clear and simple argument from @MarietjeSchaake that explains why AI regulation must *not* be left to industry to define.

Marietje Schaake @MarietjeSchaake

3 years ago

We need to keep CEOs away from AI regulation ↘️ on.ft.com/43L88ZH

13 78 242 83K 33