When GPT-3 first came out, I expected that people would use it as a sort of “common-sense reasoning module”. That is, if you want to process or generate information in some way, then you can give GPT-3 a relevant prompt, and repeatedly apply it to a bunch of different inputs to generate corresponding outputs. After GPT-3 came out, I had expected that people would end up constructing a whole bunch of such modules and wire them together to create big advanced reasoning machines. However, this doesn’t seem to have panned out; you don’t see much discussion about building LLM-based apps.
Why not? I assume that there must be something that goes wrong along the way, but what exactly goes wrong? Seems like it has the potential to teach us a lot about LLMs.
They are. GPT3 doesn’t have a lot of common sense. But, language models that large have lots of general intelligence due to their size, and are an incredible basis for doing stuff, if trained on the task at hand. eg (mostly non-vetted):
https://metaphor.systems/ (which I used to find everything in this list besides the first three items; note that you can drop any of these links into metaphor and walk the semantic relatedness web near these sites! I probably didn’t even have to paste this big list, but y’all consider link clicks a trivial inconvenience, turn up your clickthrough-and-close rate to match your usage of mindless recommenders and retake your agency over what websites you use! or something! uh anyway)
https://github.com/features/copilot (which you know about, of course)
https://summarize.tech/ (also gpt3)
general research ai tools:
https://www.semanticscholar.org/ - the paper recommender is wonderful, add your favorite safety papers to your feeds! my strongest recommendation on this list besides metaphor.
https://iris.ai/ - looks very cool but kinda expensive; probably not even available to individual researchers outside institutions
https://scite.ai/home—looks like a general related-work finding tool like semanticscholar, may have some tastefully chosen small ML models like “does this citation support or contrast?”. [costs. $16/mo individual. personally, that means SKIP]
https://www.scinapse.io/ looks cool, found via metaphor.systems semantic search seeded with semantic scholar
https://scholarmagic.com/ looks cool, claims to have cite-as-you-write tool, I wonder how it compares to galactica
https://www.onikle.io/ looks cool, claims to compete with semanticscholar, just try semanticscholar though lol,
https://www.resolute.ai/ looks cool but not free
https://www.causaly.com/ bio papers only and also not free I think
https://www.hebbia.ai/ not sure if this is for science or internal tools for companies or what
https://www.academiclabs.com/ looks xpensive
https://consensus.app/search/ tries to summarize, doesn’t do as well as the classic and locally crafted.. …https://elicit.org/! which is incredible and is the only one of these things I actually use already besides semanticscholar
https://ohmytofu.ai/ general site recommender based on contextual relevance, seems similar to metaphor in that respect[edit: ohmytofu appears nonfunctional]
https://www.albera.com/ suggests plausible research trees to learn about a subject [edit: tried! sorta works. feels like just another elicit prompt]
https://www.scholarcy.com/ paper summarizer
https://www.genei.io/ another paper summarizer, this one is actually gpt3, like you asked for
https://www.prophy.science/ not sure if this uses real ai or not but it looks maybe cool
https://www.scinapse.io/ who knows if this one is any good
https://brevi.app/ another research summarizer
https://keenious.com/ yet another paper recommender
https://www.spellbook.legal/ (custom trained?)
https://www.advocat.ai/ (might be gpt3?)
https://skritswap.com/ (looks meh)
https://www.puzzlelabs.ai/ (this one looks mildly cooler)
https://loio.com/ (ai powered contract linter)
https://www.uncoverlegal.com/ (legal semantic search)
https://www.dellalegal.com/ (contract review)
https://www.lexcheck.com/ (contract review and editing)
https://www.legartis.ai/ (contract review and editing)
https://zuva.ai/ (some blend of the above)
https://www.donna.legal/ (contract linter)
https://motionize.io/ (yet another linter addon for word)
https://www.amplified.ai/ patent recommender
not ai, but came up and looks cool: https://lawya.com/
chem & bio (cell culture go foom!) - I’m even less qualified to evaluate most of these than the legal text stuff:
https://braininterpreter.com/?q=thinking out loud ← wtf!
https://www.benevolent.com/research (actual ai bio lab)
https://atomic.ai/(lab, no public usability)
valence discovery goes here, as do deepmind and standford medai
https://www.aktana.com/ ← warning, this one looks like a manipulation tool
this one is specifically military; I’m sure that it, and many others like it, will detect this comment and categorize it somewhere: https://primer.ai/public-sector/ai-in-warfare-a-race-the-u-s-cant-afford-to-lose/
https://www.semion.io/ ← amazing looking paper relationships tool, but not actually based on deep learning
https://flywire.ai/ actually a game, not an ai
https://www.aicrowd.com/ just an ai research competition site
https://www.journalmap.org/ GIS + paper discovery? no ai tho
https://tools.kausalflow.com/ this is a list of tools made by people who like lists almost as much as I do. similar list to the stuff you find browsing my profile here—big list of tools, mildly curated but significant shopping still remains.
and of course, my purpose in sharing these link floods is to give people seeds to find stuff on the webbernet. you asked for how ai has been productive; the answer is, it’s a bit of a mess, but here’s a big list of starting points to comment on. if anyone browses through these, please share which ones were worth more than a few seconds to your intuition—I spent an hour and a half on this list and barely skimmed any of them!
Scott here from spellbook.legal (mentioned above)!
We are finding LLMs do be incredibly powerful tools for legal drafting & review, mind-blowingly good. It is a whole new way of thinking as a programmer though: results are non-deterministic! Chaining together non-deterministic queries is much more of an art than science. I think it will take the software engineering profession a long time to get comfortable with that. It really requires tinkering at scale, but not necessarily formal methods.
I also think there is a perception that GPT-3 is “too easy” and you have to “learn to do things from first principles first”. I really disagree with that, and I wrote about that fallacy here.
One last point: GPT-3 has improved dramatically over the past 2 years. It’s not the same product it was when it launched. I don’t think many people have caught on to the level of improvement yet.
I thought I might catch the eyes of some of the folks I was mentioning, heh. I’m curious which notification system you use to find mentions of your work!
also, welcome to the safety nerds site, enjoy your stay, don’t destroy the world with stark differences in effective agency, and let’s end suffering using advanced technology! :)
Your link goes to a private page, I’m afraid.
Interesting list, I had no idea there was so much.
I’d love to hear which tools from the research section you end up using! My favorites are metaphor and semantic scholar at the moment. copilot is also great for doing less typing, although it made me a mistake that I missed in some important code and I am a bit sketched out about it now.
Another example: Notion, the popular wiki/information management tool, just announced an AI-powered writing assistant. Now, they haven’t announced specifically that it’s using a LLM, but if you look at the demo, it’s hard to imagine what else it could be.
to be honest I’m slightly confused about your phrasing; it looks like they demonstrate the output of a language model on the page, and so the only question left is whether it’s transformers or some swanky high speed RWKV thing or other
I wasn’t aware of RWKV until you mentioned it. Fair enough. It’s possible that they’re using that instead of a LLM.
no I mean, that would still be an LLM. just not a transformer-based one. to not be an LLM you have to train it on significantly less text data, I think. maybe I would also count training on sufficiently much other modality data. by its behavior we can know that it could only possibly be an LLM, there could be no other AI algorithm that outputs text like that without satisfying the constraint of “is LLM”.
Oh, I guess I misunderstood what you were saying. Yes, I agree that nothing else produces output like that. I was just pointing out that Notion haven’t come out and explicitly stated what, specifically, they’re using to do this.
yeah could be any LLM. It does feel like an ungrounded generative model like most LLMs right now, but maybe it’s some swanky new physical model based thing, you never know.
There’s so many that I’m having trouble choosing just one. Can anyone recommend one for bioinformatics research? I would like something to help with hypothesis discovery, but am hoping to discover something that I currently don’t know about.
semanticscholar has been amazing, and I feel like I am often recommending new papers to people who haven’t encountered them yet thanks to its feeds; the way you use them is by adding a paper to your library, which requires an account, but it only takes a few papers before you start getting ai recommendations. if you try just one, it’s my recommendation. I’ve tried a few paper navigation tools, and my favorite so far is actually manually walking the citation graph on semanticscholar, followed by browsing its new-papers feeds.
I also have been absolutely blown away by metaphor. I’d definitely recommend trying metaphor for your paper search. it can’t do everything but it provides an incredible component and is probably the most general tool I’ve recommended here.
if you find semanticscholar and metaphor disappointing is when I’d suggest you start trying a bunch of these tools in quick succession; set a goal of a kind of discovery you’ve had before that you’d like to have again, and see if the tool can replicate it. There are a lot of really cool papers, and that’s how I find the coolest crazy-advanced-bio-whatever stuff so far; metaphor might be going to replace semanticscholar but ultimately neither are as strong as iris or causaly, afaict.
that said—I suspect that the most advanced bio tool on this list is advanced enough to make a night-and-day difference in your research throughput, and that opening all the bio links and setting a ten minute timer to close all but three would really give you some solid candidates. if you describe what you’re looking for further, I can try filtering further.
also, for baseline, I tossed your comment into metaphor with some prompt engineering; here are the results: (<icon loadingspinner/>, manually...)
foss or freeware:
https://het.io/explore/ my score: ++++++ this looks very cool! doesn’t look to be modern deep learning but rather a fairly dense plain knowledge graph
https://biokeanos.com/search my score: ++
https://scite.ai/ my score: ++ seems like a maybe interesting addition to semanticscholar but not that cool on its own
https://pharos.nih.gov/ my score: +
https://www.bioz.com/ my score: +
https://iris.ai/ my score: ++++++ looks very very cool but kinda expensive
https://academicsequitur.com/ my score: + seems like a crappy semantic scholar to me
$$$ (no price given)+:
https://www.causaly.com/ my score: ++++++++ maybe the coolest one on this list but no price given, probably a LOT
https://epistemic.ai/ my score: +++
https://www.biorelate.com/ my score: ++
https://abzu.ai/ my score: + - they have target identification
https://www.pharm.ai/ (protein binding estimate and such) my score: +
research lab focused on the topic of bio hypothesis discovery: https://discoverylab.ai/
not available yet but whoa cool:
https://www.asimov.com/ ← this is probably the most ambitious project on here, though you can’t use it right now
wat, collective behavior aggregation thing but I’m not sure if it’s good or not, or, what:
https://start.polyplexus.com/ my score: ++++
https://unanimous.ai/swarm/ real time voting system for incrementalized communication? seems like it could be prone to group bias tho
misc foss tools that were not what you seek, unrelated but cool:
https://www.h1st.ai/tutorials huh interesting transparency tool
Do you know of any AI tools where I can input a table of labeled genetic data and get out an interesting hypothesis? If nothing like that exists, I should probably make one myself.
I don’t know of one. Here’s what I found looking on semanticscholar and metaphor for
ten minutesan hour or two of diffuse-focus multitasking:
tool you could use to build it: https://genoml.com/
tool you could use to build it (contains pretrained models of questionable but possible relevance) https://torchdrug.ai/
pretrained models for genomics http://kipoi.org/about/
personalized clinical phenotyping review paper, cited by some interesting looking papers, cites some interesting looking papers. may be a useful node on the research graph, you’ll want to spider manually from here https://www.semanticscholar.org/paper/Personalized-Clinical-Phenotyping-through-Systems-Cesario-D’Oria/99aa941bb47243777731c92e3583b4e78953938b
another review of epigenetics-ml studies https://www.semanticscholar.org/paper/Artificial-Intelligence-in-Epigenetic-Studies%3A-on-Brasil-Neves/7a4589ab51c9248d05567255dad3c2acd927a27c
initial metaphor.systems query by just copying your comment didn’t do much of use for me.
also tried browsing semanticscholar from scratch using semanticscholar search, and didn’t find anything even close.
also checked my saved papers and followups to see if I had favorited anything near that—nope.
from memory, the openbioml folks have been talking about bio language models. there’s been work on text to genome or genome to text. perhaps something in that domain.
some mildly interesting results from tossing the first promising result, genoml, into m.s as a similarity search. manually filtered vaguely-cool-lookin results of projects—these seem like low quality results to me, but I’m not sure:
vaporware/sign up for access/expensive/dead startup/other misses: http://20n.com/ https://tracked.bio/ https://www.brainome.ai/ https://www.solvergen.com/
off topic but cool and funky https://www.openml.org/
almost! https://openbioml.org/ (good result, already knew of them, might be relevant to what you seek)
so, okay, let’s try “deep learning on genomics data to do causal discovery of hypotheses for protein and disease function and etiology” (search link):
vaporware/sign up for access/expensive/dead startup/other misses https://latentsci.com/
miss but interesting https://qdata.github.io/qdata-page/categories/AIbiomed/
but through it I found a paper citing it: https://www.semanticscholar.org/paper/Artificial-Intelligence-in-the-Healthcare-System%3A-Lorkowski-Grzegorowska/b97af50b0e117b272785e55a8b83067c296dfbd5
and its cited papers are interesting
still seems like https://www.causaly.com/ is your best paid option
tried semanticscholar search for “artificial intelligence in healthcare”, found some interesting results
hmm what about doing a relatedness semantic search on metaphor with a “but for genomics” prompt. ooh interesting results here
found via “artificial intelligence genomics” on semanticscholar
interesting neurosymbolic hybrid thing for explainable genomics something or other: https://www.semanticscholar.org/paper/Artificial-Intelligence-in-Biological-Modelling-Fages/f1e286a353c1b628e8613b940a222525fa6c59d8
cited by this, a very interesting lookin abstract about modern gofai for dynamical systems in biology and explainable deep learning https://www.semanticscholar.org/paper/Learning-any-memory-less-discrete-semantics-for-by-Ribeiro-Folschette/209288da6076e5a828172a4e8ca20c38ba1aeb22
misc paper on biodefense using ai https://www.semanticscholar.org/paper/Big-Data-and-Artificial-Intelligence-for-A-Approach-Valdivia-Granda/a26e064d18d49705170b31cc06cd48b21bf005d5
crop breeding with ai https://www.semanticscholar.org/paper/Applications-of-Artificial-Intelligence-in-Breeding-Khan-Wang/3b97987c4b831b0f7533aab4197c594a6de1c9d9
Thanks! I know this is super late, but this has really improved my work productivity. I really appreciate you taking the time to help.
For what it’s worth, Causaly is a disappointment. No strong LLM integration means it really struggles to compete some of the other products out there.
I don’t know about big reasoning machines, but I’ve heard a lot of rumors about LLMs being integrated into an extremely wide variety of extant ML systems that were already commercially viable on their own. It seems pretty intuitive to me that LLMs can provide some very good layers to support other systems. What have people heard about that?
GPT-3 was announced less than two and a half years ago. I don’t think it’s reasonable to assume that the market has fully absorbed its capabilities yet.
I would just have expected at least an explosion in basic demo projects that use GPT-3 for reasoning. A skilled programmer can usually code up something simple over a weekend or two, even if it is too unstable and incomplete to be economically viable. But instead there seems to just be… almost nothing.
There is already github copilot, and clones.
There is an explosion of other llms.
What do you expect? The system was never intended to be usable commercially, and it has several problems. Many of it’s answers are wrong, often enough you can’t use it to automate most jobs. And it can unpredictably emit language embarrassing to the company running it, from profanity to racist and bigoted speech, and there is no known way to guarantee it will never do that.