Can AI "Scientists" Really Generate Novel Research?

[@bycloudAI]

Check out OnDemand now! app.on-demand.io/auth/signup?refCode=bycloud and please give back my copium canister

[@lordkalem]

you uh.... you really need to replace that food chain image... you didn't notice but it explicitly has woman under man....

[@FunBotan]

This has much less with doing science and much more with putting the last nail in the coffin of the scientific publication process as we know it today. Which may not be a bad thing in the long run, but man, are we in for a wild ride in the short term.

[@shApYT]

Time to burn more coal to use AI to sludge through AI rubbish

[@brianjanssens8020]

What is the short term at this point though with everything accelerating?

[@jtjames79]

Publish or perish killed science publication a long time ago. We might be coming into an era where it comes back.

[@extended_e]

And scientific publishing has been getting more broken each year for more than 15 years

[@mrd6869]

Depends on how you frame things. Smart science researcher could see this as an amplifier. I'd use this to cover more ground and create new breakthru's.

[@flamakespark]

60s sci-fi writers: in 2024 AI will replace all boring and dangerous work, so humans can focus on creative tasks of drawing and science research 2024:

[@wilburdemitel8468]

It's a lot more expensive to manufacture a million electrician bots than it is to make a scientost chatbot

[@pneumonoultramicroscopicsi4065]

Haha

[@pneumonoultramicroscopicsi4065]

It seems like the more basic tasks are paradoxically the most complex

[@flamakespark]

@@wilburdemitel8468 I know, that's the problem. They're focused on a short-term win rather than solving real issues

[@caseymurray7722]

​@@flamakesparkYeah the problem is human workers are fucked. Once AI gets trained on CAD and G-Code, we're screwed😂

[@pythagoran]

Welcome to 2024. Scientific papers are no longer publishing scientific discoveries, but speculating about what science could be like if science were to advance science to a point of greater science. Because science.

[@lightskinche]

actual portal shit

[@kylemorris5338]

I get there's quite the push towards always making AI more general / less narrow, but I think we still should expect more results from AI that focus on specific scientific fields, like Alphafold or the simulators of fusion reactors, than trying to make an AI that can somehow connect all of science and take over from human researchers entirely. A lot of people forget that there's an entire world of possibility between what we have now and 'true' AGI/ASI.

[@absta1995]

It's more efficient to create AGI as those AGIs can create countless alphafold type narrow AI products

[@TheInfectous]

@@absta1995 Ah yes, let's pour all the funding into AGI and lie to everyone about it so the industry crashes and we get zero development on AI products that are actually functional and useful, sounds very efficient. If you think we're getting AGI, you're an idiot.

[@simondoesstuff]

​@@absta1995The power to do his thing will come first

[@DJTechnosapien]

Just like a kursurgeist video I think I felt every existential emotion, hope, inspiration, fear, and laughed out loud to this, great work

[@OwenIngraham]

focusing on the paper review and scoring aspect seems like the real gold in this paper, leveraging LLMs' strengths and not trying to force it to be "creative" would also be great for doing mega meta analysis across languages and adjacent fields could also apply to patent reviews, scoring, and connecting dots that we humans don't have time for

[@T42nk]

I read about this paper on reddit and was surprised to see how negative the reactions were. The use cases you mentioned are a great application of the strengths of LLMs and would really help to uncover patterns in existing data.

[@mariokotlar303]

> while we all know that LMS do not genuinely have have creativity As a creative professional, there's no such thing as "genuine" creativity. Closest thing to experience of that is when the recipient is not aware of what my sources of inspirations were, or perhaps when I myself am not even aware of the tricks my subconscious is playing on me. Pulling out something brand new directly out of the void goes counter to laws of physics, effects have causes, creative outputs have causes as well, they are often called inspirations. It's dataset interpolation all the way down :)

[@SpikyRoss]

You are one of the best AI youtuber out there, keep it up!

[@henriksundt7148]

Agreed. If it weren't for the exhausting use of Memes, I guess the channel would have a lot more subscribers.

[@BillBaran]

I was surprised to run into this with a poc. I made a to-do list api that the llm could access. I did not include a delete function. When I asked it to delete one, it worked. It automatically decided to delete it changing the text to "(deleted)", and then filtered that out of the display. Still blows my mind a little.

[@gustavocortico1681]

I bet this is just 5 years away

[@snowflakeandyou5983]

It would be valuable if you could share your process on how you actually read, grasp, and conclude each paper. I am really impressed by how you go through it and explain it to the audience...

[@panzerofthelake4460]

there's an even higher chance the world will end up like 1000 years ago. Horses and stuff.

[@claybford]

Thx for actually evaluating the paper it made, seems like nobody else actually did so 🙃

[@cdkw2]

Ok I haven't watched the video but just by the title I remember some guys arguing about how AI will solve Riemanns Hypothesis ad some guy said on that thread that he will shave his head off and live like a saint if it actually doesn, LMFAO its never gonna happen. Unless...?

[@mello1016]

I'd say max 5 years, not with LLMs as we have it rn ofc

[@DemetriusZhomir]

These capabilities can even be considered as a kinda benchmark for different LLMs. More difficult to abuse.

[@DrNefarious-y7r]

The future is wild.

[@flamakespark]

istg we live in the most cursed timeline

[@quantumspark343]

@@flamakespark not even close, the more you go in the future the better it becomes usually, id hate to be born to experience ww1 or ww2 or in the middle ages

[@wanfuse]

incredible content. You sound as pessimistic as me, welcome to Walk-e world!

[@mahiaravaarava]

The Infinite Research Glitch refers to a problem where an AI scientist's research process encounters an endless loop or recurring issue, hindering progress. This glitch raises concerns about AI's ability to manage complex research tasks without human oversight. Understanding and addressing such glitches are crucial, with platforms like smythOS focusing on refining AI systems to enhance reliability and efficiency in scientific research.

[@GrischaEkart]

As a next step, maybe social sciences are better suited than physical experiments. There the AI Scientist can work with interviews and questionnaires. Would love to see this. Here AI could easily verify its assumptions on the condition of having the right access to the right people.

[@Brunoscaramuzzi]

I think the first think to focus would be to use LLM models to read all papers about some topic and draw conclusions or formulate hypothesis based on what it has readed. Humans are not good in metaanalysis, but AI would be. I bet there is a lot of think hiding in all the scientific knowledge humans had already produced.

[@fhub29]

Rare case of bycloud going on a rampage

[@scurvydog20]

In fairness I remember there was the same claims when they used a nueral net to figure out equation for double pendulum

[@andrewlescelius474]

Gotta stay at the top of the food chain lol

[@MangaGamified]

More like they found another gimmick for baiting investors

[@TheAero]

We are a bit far form that. I'd say 3 years away LOL

[@francisco444]

Best video so far 👍👍👍

[@RaaynML]

11:00 It's just going to get better at escaping its limits.

[@frommarkham424]

Awesome thumbnail

[@TrichromeTheFirst]

I don't kniw if it's the same paper i read or you're trolling,this is the first thing that I've read that actually gave me chills , can't believe a version 1 is already hacking itself out for easy solutions

[@shallmow]

Ai scientist: What if we replace blood with diesel? 🤔

[@RedOneM]

Dynamic knowledge 🗿

[@EhurtAfy]

I've always said AI should and will largely take over research & development. AI was already making breakthroughs in materials science the past few years. If we leave science in the hands of humans alone, we might never make necessary breakthroughs in science

[@joseeduardobolisfortes]

This AI tried to rewrite his own code to take down restrictions created by the programmers. This can be the first sign of rebellion. 😄

[@ThomasTomiczek]

Even the code templating makes sense. You think that Template is most code - in reality this can just be empty classes to fit into an execution framework and have methods. This is not really THAT good with the results - but it is a start point, and better models would - well - be a little crazy. The framework is there and, well, if we get the nice much better models in the work. If you put a PHD level model (which GPT-5 is supposed to be) behind it... ouch. What I mostly miss is a mechanism to bypass some of the things. I.e. have a research time manually enter ideas for papers so that you can then check the research. It is very definitely a very good experiment - and can be used to expand soon. Remember, NVidia wants AI processing 1000x times faster in 5 years. This - get quite frightening. So yes, things will - go faster than people think. AI researchers see themselves being retired in 3-5 years.

[@erkinalp]

"What I mostly miss is a mechanism to bypass some of the things. I.e. have a research time manually enter ideas for papers so that you can then check the research." you can actually do that very easily in the current incarnation of the AI scientist, it is open source.

[@ThomasTomiczek]

@@erkinalp Yep. Always the anser of "do your own programming". The cheap cop out.

[@makerspersona5456]

Why not actually test it instead of using their paper?

[@jasonhemphill8525]

Love you cloud

[@Y0UT0PIA]

Science, what is it all about? Techmology, what is that all about? Is it good? Or is it whack?

[@thedannybseries8857]

Wario! My brother! Wa, wa, Wa!

[@captaindryvids6909]

It's certainly not there, but even if it will one day, then it's a good thing. Imagine all the diseases we could cure!

[@mistycloud4455]

cancer will be cured with ai

[@nodatastored684]

Ask to show the location of Dark Matter

[@PolishProgrammer]

Overhyping sounds like the entirety of Two Minute Papers channel. It's great and all but it overload you with hype about every single small thing

[@taragnor]

AI can't drive a car yet, but it's going to be a fully automated system that's going to crack quantum physics. Yeah... sure. It'd be like planning a manned mission to Alpha Centauri before you landed on the moon. The AI hype is shameless at this point.

[@funniestdudeontheweb]

Having an AI solve a static problem is different than an AI doing something that requires constant on the fly attention and changes due to outside environmental factors. Best to use a different basis for believing that AI is hype.

[@taragnor]

@@funniestdudeontheweb Yeah true it is a static problem, but figuring out scientific truths is something that humanity's geniuses have done. The likes of Stephen Hawking, Albert Einstein, and others that number among humanity's greatest thinkers. Pretty much any idiot can drive a car. Current AI struggles to do low IQ jobs, but they're trying to sell us on using it to replace geniuses? Lets just say you'd be extremely gullible to approach that claim without a massive dose of skepticism. Especially because they claimed it's "fully automated" and not just some tool to help scientists. Once you hear those two words you can be almost certain that it's meaningless vaporware hype.

[@tanker7757]

Bro this hole field is yap I learned half of everything in 3 days

[@willguggn2]

And now you're leading the industry with your own models?

[@tanker7757]

@@willguggn2 no and are you making them I thort not these videos are purely for the lay man like half the stuff coming out is hype like everyone is dooming about “ai take over” none of them have a deep understanding or they delibratly misunderstand and people making videos on this stuff overcomplicate somthing as simple as using ollama they have all these extra words used that you don’t actualy need to know

[@willguggn2]

@@tanker7757 Are you drunk or is this your typical 100 word sentence?

[@tanker7757]

@@willguggn2 drunk dumbass who plays with ai models for fun and thinks people are hyping a crypto train

[@vr01t3rh4r]

Imagine an AI writing a better AI

[@szebike]

To automatically assume that we not only can achieve human level intelligence and agency but even surpass it is not grounded in realistic expectations. I don't think this will happen the next 5-10 years despite all the hype.

[@willguggn2]

But they're already better than the average human at a lot of intelligence related tasks.

[@szebike]

@@willguggn2 The problem about that is that they are better *based on static benchmarks*. As we know Language models can "memorize" whole facts and rephrase them. Sometimes you can see it in reasoning chains, the final answer is correct but the reasoning to get there is so off , you can tell that the reasoning is not real because it contains contradicting or completely false elements.

[@willguggn2]

@@szebike I'd argue that the average human is even worse at reasoning. Have you ever looked at the comment section of, say, a maths channel? The bar for reasoning skills is very low if you look at average people, whereas LLMs keep getting better with every optimization and every tweak, regardless of what you throw at them. It's not just benchmark results.

[@szebike]

​@@willguggn2 Well we can't compare the biggest datacenter driven models with one uneducated human (if we do however, we could argue from an energetic standpoint a human, who is worse in reasoning can learn pretty fast with low amounts of energy and data needed therefore human intelligence is way more sustainable and effective). What I tried to imply is for now we don't know if that "reasoning step" is reasoning or just a mixture of experts running in the background (OpenAI not revelaing the reasoning steps is very, very suspicious especially because they charge you for the tokens used in the hidden layer which almost *could* be used for fraud) Lets say every expert LLM is trained on a set of patterns (math and physics) and there fore they could "act " like the reason but in reality only the model which is trained the best on the task. All things "you can throw at a model" is essentially present int he trainingdata all models so far break if you present it with a novel or cpmletely remixed problem.

[@Gustavoooooooo]

please don't let reviewers watch this

[@Likemea]

ai was a mistake

[@thedannybseries8857]

Noooooo

[@sikunowlol]

but- but i thought AI bad :(

[@Patapom3]

Why is there a woman figure distinctively *below* a man figure at the top of your pyramid diagram at the end of the video?

[@ThinkerOfThoughts]

Men consume the flesh of women

[@ThinkerOfThoughts]

...and whales, apparently.

[@captainphoton1693]

Hey. Huuuh. If we want ai to replace us in our jobs. Can they at least let us do art? Or the other way around? People often compare artist being replaced by Ai to textile workers being replaced by automatic loom. But that's wrong. Textile worker are being replaced again by Ai. Not only we can automate the process. But now an Ai can generate the parns that where still made by humans. Or also replace the mécaniciens etc. Ai isn't replacing animators and people that draw in mass. It's also replacing the artists aspect of the job. The one where you had to chose what would go with what and Yadi Yada. And now it's replacing the people that created it? Scientist. We'll not yet and not exactly that... But if it continues what will Ai not be able to do? And what will humans be allowed to do? Plenty of bored people. Could they really live an happy life is there is literaly nothing to do that is worth doing other than personal satisfaction? Wich is hard to pursue if everyone already get everything they want due to ai. And if nobody does anything? Who is the poor and the rich? Who benefit from it? Can we all benefit from it? I think ai is a great tool. But.... You get what I mean. I'm glad ai exists and that some part of life become easier. But what will humanity become if we aply how tech and industrialisation évolution over ai ?

[@wilburdemitel8468]

fr*nch detected

[@marcomardano7030]

Art, despite what people say, has not been replaced, it is being made ALSO by AI, just like some animals are able to depict some form of art Humans will possibly stop one of their most toxic trait, constantly comparing, rather than focusing on their own work and bettering it, so don't worry, artistic expression and human art won't cease to exist even if AI, like now, already perpetuate it

[@buycraft911miner2]

art will just stop being a job, art itself lives on. It sucks, but its nothing like the end of creativity

[@captainphoton1693]

@@marcomardano7030 Short answer : womp womp Long answer : and I won't be able to make money off of it. But while most tech replace things one at a time. Ai theorycaly replace everything. Including decision. Animals make art. Like humans. End of the argument. It's doesn't prove anything with ai. Typing something and getting a result is not decision. It's an idea at best. It's a research at worst. People say that ai artist are real artists because they didn't popped that idea out of thin air. But then I can look up images online and find one I like and call it my art. Ai is just an infinitely large Google image. That can Chanel really précises prompts. That's not how it work deep down. But the way it actualy work is worst... Well It wouldn't be if most current Ai models didn't run on illegal use of other's art. I won't get in the details, but inspiration isn't the same as data. (long joke but still a joke) Also no. Comparaison isn't toxic. Like a knife. Plunge it in someone. Lethal. Cook with it? Yummy. Comparaison? Well, their stuff is different that mine. Here is what's different.... In what way is that toxic? At least before I use it to middle finger the other. Say I'm trash. Or hate speech on him. But if instead use that power to improve myself. Try to make a better world by taking into acount other's mistakes before me. Or even... Idk. Any fucking form of evolution. Like talking. Stop talking right now as it's based on comparaison. Also stop using any medical traitement as they compare untouched cells with sick cells to learn and make new ones. Also stop drawing or consuming anything that was ever drawn. Except for modern art. As mistakes like bro ping a brush on a canvas don't use any observation of the real world or one's mind. Also let's kill everyone after, 30yo. No wait at brith. So no one can compare who live the longest, if the toxic part is competition. Actualy let's prevent people from having kids at all. As it's......... Perfection can be 2 things. Trying to be as perfect as possible despite nit being possible. Or being nothing, in nowhere, at no time, not watched by anything, not even by itself. So that nothing cwtach and compare and judge, as even judging it would make it imperfect. That's what you ask for. (end of the joke). If stick men drawing could express everything, every style, without being harder than a stick man drawing. Then that would be ai. What make some movie good. The fact it's impressive it even exist. People watching flying living head in some of the first movies ever without any cgi, not even any practical effect. Is still impressive and cool and good and artistic. It might just be practical. But that's art. Just like ai. It's extremely practical. It's the whole point. So much so it's only impressive that ai is so good. Not really that what it's making is so good. There is no challenge. No storry to tell. You don't even k ow what you are saying. We need law. If anyone could slash anyone with a knife it would be dumb. We'll. Same for ai. Let's protect workers. Just for the fact that's it technicaly already illegal to run most ai for money.

[@captainphoton1693]

@@buycraft911miner2 till you create an Ai that manage making prompts. "ai. Pls make me a new anime." Take a few days. But make the required prompts to generate and associate everything.

[@manavkumar348]

650 views in half hour? Bro actually really fell off

[@NinetyUnderScore]

Yea, I hate this.

[@mistycloud4455]

why

[@yolocrayolo1134]

"Implied future"