I do not believe human-level AI (artificial superintelligence, or the commonest sense of
#AGI
) is close at hand. AI has made breakthroughs, but the claim of AGI by 2030 is as laughable as claims of AGI by 1980 are in retrospect. Look how similar the rhetoric was in
@LIFE
in 1970!
I’ve kept quiet on the
@OpenAI
fiasco, since I also don’t know what’s going on, 🤷 but I can’t possibly support today’s interim CEO—the below in a thread on “50/50 everyone gets paperclipped & dies”—or a residue board that believes in these EA-infused fantasy lands. HT
@vkhosla
.
@BarbettiJames
@ApriiSR
@BellaRudd1
The Nazis were very evil, but I'd rather the actual literal Nazis take over the world forever than flip a coin on the end of all value.
“The fact that [transformer neural nets] model language is probably one of the biggest discoveries in history. That you can learn language by just predicting the next word with a Markov chain—that’s just shocking to me,” Mikhail Belkin says. By
@strwbilly
.
This paper gives some really nice insights and mathematical depth to what had previously (for us) been “the mystery of squared distance” in revealing the representation of parse trees in deep contextual representations (BERT, ELMo, etc.). Great to read!
This is truly an opinion piece. Not even a cursory attempt is made to check easily refutable claims (“they may well predict, incorrectly”). Melodramatic claims of inadequacy are made not of specific current models but any possible machine learning approach
Artificial Intelligence Definitions: This (northern) summer, I spent more time than I’d like to admit coming up with a handout defining key terms in AI in 1 page, trying to be informative and suitable for non-specialists – let me know if you like them!
UC Berkeley to limit Computer Science degrees even harder. Even with intensely hard lower div courses, a 3.3 lower div GPA requirement to declare CS, they could still only weed down +2,000 intro CS courses to 800 a semester. A sad state of education.
But most AI people work in the quiet middle: We see huge benefits from people using AI in healthcare, education, …, and we see serious AI risks & harms but believe we can minimize them with careful engineering & regulation, just as happened with electricity, cars, planes, ….
Dear
@emilymbender
—and
@Abebab
—you need to keep “reminding” people of your viewpoint because it is not an argument that is convincing to all or a self-evident truth. It is a particular academic position, which lots of people support but a good number of others disagree with. 1/8
Yes, exactly this. I wish we didn't need to keep reminding people, and
@Abebab
is commendable for being gentle about it!
For the long form of this argument, see Bender &
@alkoller
2020:
Reflecting again on how knowing all the architecture & equations of the Transformer model is really of no use at all in convincingly explaining to someone how an LLM like ChatGPT can write paragraphs of lucid text in response to a prompt.
I guess I’m saying “Beware reductionism”.
LLMs and other generative AI are enormously powerful, because they soak up, abstract, and can mashup the work of millions of humans. But they are only a bit more intelligent than an encyclopedia. Central to intelligence is the ability to learn, adapt, and act in novel situations.
People most-cited by
#AAAI
papers shows 25 years of AI history. 1990s greats: Pearl—Kautz—Weld—Selman; rise&fall: Comitzer—Sandholm—Sutton—Domingos—Tambe—Littman—Jordan—Veloso—Koller—Boutilier—Ng—Barto; 2010s neural boomers: Bengio—Sutskever—Hinton—Manning
LLMs like ChatGPT are an amazingly powerful breakthrough in AI and a transformative general purpose technology, like electricity or the internet. LLMs will reshape work and our lives this decade. They are not just a blurry photocopier or an extruder of meaningless word sequences.
👇 Honestly, this thread is 80% wrong. This is treating science as like front-end frameworks. Yes, if you’re a front-end developer who only knows 3-years-old JavaScript frameworks, then you’ll have trouble getting a gig. But that’s not what we’re teaching students 1/7
AI has 2 loud groups: “AI Safety” builds hype by evoking existential risks from AI to distract from the real harms, while developing AI at full speed; “AI Ethics” sees AI faults & dangers everywhere—building their brand of “criti-hype”, claiming the wise path is to not use AI.
I’d long wondered whether physicists were making good use of all those supercomputer clusters or just using really inefficient algorithms – it looks like some answers might be emerging. Simple AI shortcuts speed up simulations by billions of times | AAAS
COVID-19 and AI: A Virtual Conference – Stanford’s Human-Centered Artificial Intelligence Institute (HAI) presents a special 1-day online conference, live-streamed starting 9am Pacific time, tomorrow, Wed April 1 (no joke!)
🏅 To me, this feels more like the kind of neural model interpretability research we should be doing than much of the recent work on interpretability of transformer models.
Emergence in LLMs is a mystery. Emergence in physics is linked to phase transitions. We identify a phase transition between semantic and positional learning in a toy model of dot-product attention. Very excited about this one!
I agree. Too many PhD guidebooks recommend choosing a senior prof as advisor. Going with a new faculty may lead to a few rough edges, but most students do very well, benefiting from the top current knowledge, enthusiasm, time commitment, and aligned goals of the new faculty.
If you're seeking an advisor in NLP, consider exploring options with recently appointed faculty. While established senior folks are well-known, junior folks shine as rising stars. In fact, many successful PhDs are the first few students of their advisor. Don't miss the chance!
Pleased to be promoting Human-Centered Artificial Intelligence. Our focus areas: developing AI inspired by human intelligence; guiding, forecasting, and studying the impact of AI on human society; designing AI applications that enhance human capabilities.
The amazing rise of reinforcement learning!
(With graph neural networks and meta-learning in hot pursuit. ConvNets? Tired.) Based on
#ICLR2021
keywords HT
@PetarV_93
My attempt at understandable but technically correct definitions for key terms in Artificial Intelligence in one page for
@StanfordHAI
. With thanks for helpful feedback from people on Twitter, I’ve revised and hopefully improved a few of the definitions:
Now that everyone is writing LLM programs, the idea of doing approximate bayesian inference by sampling along linguistic pipelines (rather than k-best, etc.) is more relevant again
The way I would improve spreadsheets is by allowing row numbering to start from any integer (including negative). Then row numbers could count what it makes sense to count. Books have done this forever via Roman numerals for front matter. Surely this isn’t so hard to do in 2021?
One of the simplest but most useful and appropriate pieces of AI regulation to adopt at the moment is to require model providers to document the training data they used. This is something that the
@EU_Commission
AI Act gets right … on p.62 of its 272 pages (!).
So when *the CTO* of OpenAI is asked if Sora was trained on YouTube videos, she says “actually I’m not sure” and refuses to discuss all further questions about the training data. Either a rather stunning level of ignorance of her own product, or a lie—pretty damning either way!
I’m happy to share the published version of our ConVIRT algorithm, appearing in
#MLHC2022
(PMLR 182). In 2020, this was a pioneering work in contrastive learning of perception by using naturally occurring paired text. Unfortunately, things took a winding path from there. 🧵👇
Just worry less about AI hype? Lots 50 years ago—and world didn’t end: “In 1970, Life Magazine, overstating [Shakey the robot’s] abilities, called it ‘the first electronic person’ and suggested that true ‘thinking’ machines would arrive in the near future“
The need for open data & benchmarks in modern ML research has led to an outpouring of
#NLProc
data creation. But
@harm_devries
,
@DBahdanau
& I suggest the low ecological validity of most of this data undermines the resulting research. Comments welcome!
GPT-4 is still 🥇, but I stare, amazed at how good “mostly open” LLMs have become in the last 6 months. On LMSYS Chatbot Arena , the best open LLMs—Mixtral-8x7b-Instruct & Yi-34B-Chat—are roughly tied vs. ChatGPT-3.5 and get ~30% wins against GPT-4-Turbo.
🧐
@GoogleAI
’s neural machine translation isn’t yet perfect This is a good example of how neural language models still go haywire, especially when training data is sparse. See the discussion in
@YejinChoinka
’s .
In Artificial Intelligence, all the good ideas are on
@arxiv
—together with many mediocre and bad ones. Students come to
@StanfordAILab
not to steal ideas but for training and community—to learn the creativity and boldness of thought that advances science.
I co-chaired the 1st
@iclr_conf
with
@AaronCourville
&
@rob_fergus
in 2013. An exciting 3 days in Arizona but it was small: no area chairs—we did all paper decisions. Now—in < 10 years—ICLR is the highest h5-rank ML conference—above NeurIPS & ICML. Wow! 😲
.
@Thom_Wolf
: “Academia is back as we saw at NeurIPS 2023. With many private and open-source labs closing the doors on publishing their results and data, academia rises again in visibility and is shining with many impactful papers in 2023 and exciting new work coming.”
Some predictions for 2024 – keeping only the more controversial ones. You certainly saw the non-controversial ones (multimodality, etc) already
1. At least 10 new unicorn companies building SOTA open foundation models in 2024
Stars are so aligned:
- a smart, small and dedicated…
OpenAI fires women on the board- (board chair who oversaw fuck up stays🤷♀️) - joining board is *Larry Summers* who once said women don’t have the same ‘intrinsic aptitude’ for STEM, and associated with Jeffrey Epstein even after he was convicted of sex offences
Early 2023 vibes: The AI Ethics crowd continues to promote a narrative of generative AI models being too biased, unreliable & dangerous to use, but, upon deployment, people love how these models give new possibilities to transform how we work, find information & amuse ourselves
Just a couple of years ago, I found it hard to believe that vision people were still all working with rectangular bounding boxes. I guess they’ve fixed that now. 🙂
Today we are releasing Mask R-CNN Benchmark: a fast and modular implementation for Faster R-CNN and Mask R-CNN written entirely in
@PyTorch
1.0. It brings up to 30% speedup compared to mmdetection during training. Check out the webcam demo!
Re-upping a piece from last year by
@hamandcheese
on LLMs and language meaning:
“I see the success of LLMs as vindicating the use theory of meaning, especially when contrasted with the failure of symbolic approaches to natural language processing.”
The enthusiasm of big generative AI/foundation model companies to claim AI existential risks and ask to be regulated by the government is the same old story of “Regulation is the friend of the incumbent”, and especially damaging for open source, argues
@bgurley
.
HT
@s_batzoglou
Excited that—after a lot of work—the
@Stanford
Institute for Human-Centered AI is launching. We’re aiming at new AI applications that augment human capabilities through both developing new AI technologies and studying and guiding the human and societal impact of AI.
Hi, we are the
@Stanford
Institute for Human-Centered Artificial Intelligence. AI has the potential to transform our world – how will we ensure it improves life for all of us? Join us in our work to explore this dream of a better future.
#StanfordHumanAI
we're starting to see top companies spend the same amount on RLHF and compute in training ChatGPT-like LLMs
for example, OpenAI hired >1000 devs to RLHF their code models
crazy—but soon companies will start spending $ hundreds of Ms or $ billions on RLHF, just as w/compute
“Open source is indisputably one of the biggest drivers of progress in software and by extension AI. But it is under existential threat from regulation that will advantage entrenched interests. We believe that open AI is vital for research, innovation, competition, and safety.”
Open source is one of the biggest drivers of progress in software - AI would be unrecognizable without it.
However, it is under existential threat from both regulation and well-funded lobby groups.
The community needs to defend it vigorously. 🧵
Heading to Seattle for
#NAACL2022
. This will be my first travel to an in-person conference in over 2 ½ years (NeurIPS2019 in Vancouver to NAACL2022 in Seattle—but not via Puget Sound)
Surprised normally rigorous
@beenwrekt
calls this blog post excellent—I’d say poorly argued. 1st argument for deep learning stalling:
@AndrewYNg
tweeting less. 🤔 I put his data in a chart—because
#infovis
. Anyway, does rate correlate with AI or Ng’s jobs?
Excellent post by
@filippie509
on saturation of the deep learning revolution. The only thing I’d add is that user-facing AI inside the big companies is already failing us at scale (recommendations, ads, engagement). More depth won't fix these problems.
Human-in-the-loop reinforcement learning—DaVinci instruct—may be the most impactful 2022 development in foundation models. What can we achieve by reinventing the AI design process to start from people’s needs?
Watch tomorrow’s
@StanfordHAI
conference 9 PST
@Simeon_Cps
@LIFE
A system that gains memories from one event, develops novel plans consistent with constraints, understands the implications of a changed environment, & reasons about new circumstances—without regular dumb goofs showing there’s no real world model & reasoning behind the curtain
“The most important thing to remember about tech doomerism in general is that it’s a form of advertising, a species of hype.”
The apocalypse isn’t coming. We must resist cynicism and fear about AI | Stephen Marche | The Guardian
If you’re interested in LLMs, RLHF, etc.,
@natolambert
has been doing a great series of interesting posts at interconnects dot ai. But I think I’m not meant to link to them these days on Twitter, right?
As a Professor of Linguistics myself, I find it a little sad that someone who while young was a profound innovator in linguistics and more is now conservatively trying to block exciting new approaches.
For more detailed critiques, I recommend my colleagues
@adelegoldberg1
and
It’s great to see Bean Machine, a new Probabilistic Programming Language (a bit like
@mcmc_stan
) built on
@PyTorch
.
But how much impact will this have? Somehow Bayesian modeling has gone from the center of AI in the 2000s decade to the margins since 2015.
1.5 MB really feels too low to me … but maybe I should read the article first or spend more time on compressing neural language models before commenting further. 🤔 [Kids store 1.5 megabytes of information to master their native language | Berkeley News]
NLP is having a moment, where LLMs have become the Swiss Army knife of almost all AI, but, nevertheless, I was only trying to give a brief history of
#NLProc
not AI.
However, I’ll go with it being wonderful and easy to read. 😊
.
@ylecun
& J Browning’s What AI Can Tell Us About Intelligence in Noēma is excellent! 👍
It clearly & dispassionately contrasts two main views on the place of symbols, as hard-coded at the outset or learned through experience, arguing well for the latter.
To use the currently trendy terminology, what we’re teaching students is ✨meta-learning✨—a strong foundation of approaches, ideas, understanding, and tools so that they will be able to quickly learn and evolve over the following decades, as science and engineering changes 2/7
I succumbed to threats and wrote my 2019–20 faculty report. Hot papers last year: Electra: Pre-training text encoders as discriminators , Stanza: A Python toolkit for many languages & Universal Dependencies v2
I can question particular classifications (SHRDLU equal to unskilled human or Grammarly at Level 3 seems generous), but:
This paper is a sensible, concrete framework for assessing progress towards AGI.
Congrats to
@Stanford
grads
@merrierm
&
@jaschasd
!
It is admirable to apply the precautionary principle, and build & deploy transformative AI technology with exceptional care. But how is that best achieved by distracting from very real AI risks by making a remote, fanciful risk of extinction from AI a global priority? 🤔
I’ve worked my whole life on AI because I believe in its incredible potential to advance science & medicine, and improve billions of people's lives. But as with any transformative technology we should apply the precautionary principle, and build & deploy it with exceptional care
I really recommend this 20 minute video. It makes really tangible the obstacles coming from a lack of AI community, teachers, and mentors—oh, and electricity—but also creative and successful ways to circumvent these obstacles
Watch legendary
@black_in_ai
DJ Hassan discuss his journey in machine learning, distributed research outside the confines of academia, & how he was influenced by BAI &
@DeepIndaba
.
@kahneman_daniel
calls it: “Clearly AI is going to win” [against human intelligence], and lots of other interesting thoughts on system noise, exponentials, and human judgments. The big remaining question is how to use AI advances to augment human lives.
I have been working on vision+language models (VLMs) for a decade.
And every few years, this community re-discovers the same lesson -- that on difficult tasks, VLMs regress to being nearly blind!
Visual content provides minor improvement to a VLM over an LLM, even when these…
Introducing Pika 1.0, the idea-to-video platform that brings your creativity to life.
Create and edit your videos with AI.
Rolling out to new users on web and discord, starting today. Sign up at
I’m on Jeremy’s side! A paper—with the same issues of the inscrutability of GPT-4—claims “GPT-4 Can’t Reason” by examining “21 diverse reasoning problems”. Showing that a lets-think-step-by-step prompt is enough for it to solve the first 3 seems a worthy contribution for a tweet.
Sensible words from a sensible bloke: “The Senate hearings, I felt a bit sad. AI has potential to optimize healthcare so we can implement a better, more equitable system, but none of that was actually discussed.”—
@kchonyc
via
@sharongoldman
My 2 point plan for improving productivity and world GDP:
@Apple
,
@Google
,
@Microsoft
coordinate a change so:
• {Cmd|Ctrl}+V is Paste {Without|and Match} Formatting
• In documents with paragraphs/lines {Cmd|Ctrl}+A first selects a paragraph/line; press it twice to “Select All”.
Many computer scientists have been slow to appreciate the possibilities of neural networks – “What a waste of talent,” Alan Eustace said – but not visionary
@JeffDean
“The success of this ‘misdirected’ effort [i.e., building LLMs] has tended to support theories of meaning that explain it instead as a collective phenomenon—like Lévi-Strauss’s ‘universe made up of meanings’ or Foucault’s Archaeology of Knowledge (1969).”
I finally read
@boazbaraktcs
’s blog on DL vs Stats.
A great mind-clearing read! 👍
“Yes, that was how we thought about NNs losing out due to bias/variance in ~2000”
“Yes, pre-trained models really are different to classical stats, even if math is the same”
50 years after John McCarthy’s Turing Award lecture on The Present State of Research on AI, what are now the key issues? Join
@drfeifei
& me on Tuesday as we focus on Foundation Models, achieving accountable AI, and AI modeling physical & simulated worlds.
Looking back on the hype, VC funding, and huge genuine progress in AI, ML, and autonomous vehicles in the 2010s, I think this will come to be seen as an inflection point: Uber, After Years of Trying, Is Handing off Its Self-Driving Car Project
Huge props to the `trl` team at
@huggingface
for authoring the best content around doing all kinds of policy optimizations for LLMs.
They do it keeping accessibility at the forefront 🤗
Hope to bring some of that to 🧨 diffusers someday.
For now, enjoy
But it is shocking that next word prediction can drive learning the fine structure and meaning of human languages, profoundly so given Chomsky’s claims that dominated much of linguistics, while LLMs work less well on unnatural signals; see
@JulieKallini
:
It must take a very particular kind of blindness to not be able to see that we have made substantial steps—indeed, amazing progress—towards AI over the last decade …
10 years after deep learning's breakthrough year,
@venturebeat
spoke to AI pioneers Geoffrey Hinton, Yann LeCun and Fei-Fei Li, who say rapid progress in
#AI
will continue. But critics push back on hype, limitations and ethics/bias issues. Read more:
My “Human Language Understanding & Reasoning” in
@americanacad
’s Dædalus is a short, readable intro to language understanding and generation by computers (“artificial intelligence”).
Thousands have read it in the last 3 months … so maybe you should too?
Hey
@spacy_io
people (
@honnibal
,
@_inesmontani
), those speed comparisons on are not only outdated—as you note—but the speed for the Stanford Tokenizer is just way wrong. Time to take them down? Here are our measurements:
#NLProc
.
@SuryaGanguli
and I discuss
@StanfordHAI
’s 6 hour online conference coming on Wed Oct 7, exploring the latest in machine learning, artificial intelligence, neuroscience, psychology, and how better to meld their insights; hashtag:
#neuroHAI
. Register now.
When deep learning took off 2010–20, so few in Systems knew NNets or even matmuls. AI folk had to learn Systems to be stars like Krizhevsky &
@ilyasut
! Now, many great Systems folk can make NNets go brrr. It’s high time for AI scientists to focus on novel AI modeling ideas again!
This does show something fascinating! But not that linguists’ knowledge of language is “bunk”. Rather, what has mainly been a descriptive science—despite what Chomsky claims!—hasn’t provided the necessary insights for engineering systems that acquire and understand language use.
Happy New Year.
Class that everyone's getting into language
Weird that the folks doing it best are folks who've not really spent that time studying 'language'
Suspect this reveals underlying truth - that extant knowledge of lang, of the sort that folks like me have, was bunk
It’ll be interesting to see in a year’s time how the distribution of number of citations varies between Findings of EMNLP 2020 and regular EMNLP 2020 papers. Just randomly skimming papers as they appear on social media, a lot of the time they look equally interesting to me. 🤔
I would suggest that this thread errs by over-representing the proportion of the time in which human “reasoning” is actually anything akin to mathematical reasoning, such as the example of solving SAT instances. 1/
The impressive deep pattern recognition abilities of
#DNN
's such as
#LLM
's are sometimes confused for reasoning abilities
I can learn to guess, with high accuracy, whether a SAT instance is satisfiable or not, but this not the same as knowing how to solve SAT. Let me explain. 1/
Opening theme at
@DigEconLab
workshop:
Human-level AI is the wrong goal! We should not seek AI that does what humans do—leading to AI competing with humans, reducing their power/wages—but should look away from lamplight for AI that augments humans, increasing the value of people