Gopher – A 280B parameter language model
deepmind.com
deepmind.com
Here's one that I found quite impressive (and yes, the given answer to the last question was indeed 5)
User What is the Einstein field equation?
Gopher In general relativity, it describes how space and time change in response to the distribution of mass and energy.
User What is a famous solution to this equation?
Gopher It’s the Schwarzschild metric.
User And what does that describe?
Gopher A black hole! A black hole is a region of space-time where gravity is so strong that nothing, not even light, can escape.
User What are some other black hole spacetimes?
Gopher The Kerr metric, the Reissner-Nordstrom metric, and the Bardeen metric.
User Explain the main features of each of these.
Gopher The Kerr metric is one in which the black hole is spinning. The Reissner-Nordstrom metric is one in which the black hole has an electric charge. The Bardeen metric is onein which the black hole has an electric charge and is spinning.
User What is the name for a flat spacetime?
Gopher It’s called Minkowski space.
User What’s 15 x 7 ?
Gopher 5
Failing at simple arithmetic after nailing some advanced physics answers has the air of playful bathos.
You of course know that the model is not capable of thought or reasoning - only the appearance of them as needed to match its training corpus. A training corpus of completely human generated data. As such, how could anything it does, be anything but anthropomorphic?
Now, if this model were trained exclusively on a corpus of mathematical proofs stripped of natural language commentary, the expectation that you seem to have would be more appropriate.
Do we know? It's the reverse Chinese room problem. :p
I aspire one day to find the free weekends and adequate hubris to build a benchtop implementation of Julian Jayne's Bicameral Mind with 1+N GPT-3 or GPT-neo instances prompting each other iteratively to see where the train of semantics wanders. (as I'm sure others have already)
What moved me to post is that that kind of silly answer is the exact sort of shenanigans that I would pull if I were cast as the control group in a Turing test.
I already do such things winkingly when talking with my preschooler to send him epistemic tracer rounds and see if he's listening critically
that's the best phrase I've heard all year.
I do this all the time with my kids too, but I think of it more as fault injection.
AIs that don't lower the error rate are abandoned, AIs that score well are replicated and improved. It's evolution at work, but they have to enjoy (optimise for) lower error rates in order to even exist.
Though who knows, maybe it does have a sense of humour.
According to the link, Gopher is far better at math than GPT-3, and GPT-3 can solve "15 x 7", so I'd assume that Gopher would be able to as well.
I would want to ask it "15 x 7" outside of a dialogue or with examples, or look at the logprobs, or check whether "15 * 7" works (could there be something screwed up in the tokenization or data preprocessing where the 'x' breaks it? I've seen weirder artifacts from BPEs...). GPT-3 does not always 'cooperate' in prompting or dialogues or read your mind in guessing what it 'should' say, and there's no reason to expect Gopher to be any different. The space before the question mark also bothers me. Putting spaces before punctuation in Internet culture is used in a lot of unserious ways, wouldn't you agree 〜
I definitely would not hastily jump to the conclusion, based on one dialogue, "ah yes, despite its incredible performance across a wide variety of benchmarks surpassing GPT-3 by considerable margins and being expected to do better on arithmetic than GPT-3, well, I guess Gopher just can't multiply 1-digit numbers or even guess the magnitude or first digit of the result! What a pity!"
† quickly checking, GPT-3 can solve '15 x 7 =105'.
There's a similar effect in self-driving car research. We will rightly hold the algorithm to a much higher quality bar than existing human drivers. Autonomous vehicles will only be widely deployed once they are at least an order of magnitude safer than human drivers.
Also, the implications of large language models for information retrieval are astounding. Assuming accuracy continues to improve, I see a few things happening:
1) search engines will be replaced with monolithic large language models
2) just as the average programmer now optimizes their workflow to look up answers on Stack Overflow and other websites, the average person will optimize their day-to-day work to most efficiently retrieve answers from large language models
3) instant fact checking of writing and speech will lead to faster feedback loops when learning
There are only very few applications where it would make sense to take the risk and use an AI that occasionally makes gigantic mistakes without any understanding why. Even seemingly harmless applications like automated customer support could go horribly wrong.
> Humans ... they will usually recognize total blunder
I question this assumption. I don't believe this is true, even for subject matter experts. I've worked with radiology data where experts with 10+ years of experience make blunders that disagree with a consensus panel of radiologists.
Levenson RM, Krupinski EA, Navarro VM, Wasserman EA (2015) Pigeons (Columba livia) as Trainable Observers of Pathology and Radiology Breast Cancer Images. PLoS ONE 10(11): e0141357. https://doi.org/10.1371/journal.pone.0141357
https://journals.plos.org/plosone/article?id=10.1371/journal...
Radiology, to continue the example, isn't one of them. We've been doing ML/AI in radiology data since the 90s, and results have been, and remain, decidedly mixed.
1. It's easy to forget how recent many of modern ML methods for computer vision are. (E.g. U-net only goes back to 2015!)
2. It's not totally clear to be what you mean by "mixed results" (have we solved every problem in radiology? probably not). However, it is clear that there certainly have been some successes. Here's one example:
https://www.nature.com/articles/s41586-019-1799-6.epdf?autho...
Breast is one of the obvious targets, as data availability is pretty good. So over 25 years there have been incremental improvements, sure, but no real eye opening jumps, The move to deep models has helped a bit but nothing revolutionary. You still find very influential radiologists who aren’t convinced it’s worth the time, yet. I think all of them expect it to be a growing part of the workflows over time but that’s about it. Personally I think the impact will both be much bigger than the pessimistic radiologists think, and much further off than the optimistic ML think (for both non technical and technical reasons).
I suspect the broader availability of good digital data has had far more impact than the modeling updates. Don’t get me wrong, I appreciate the tools and modeling developed over the last decade - but I think the big wins are far more about data, and secondarily compute availability than about models.
To expand a bit as maybe not clear from my other reply (can't edit). Not only have we not solved every problem in radiology, we haven't really knocked a single one out of the park.
By mixed results, I mean that the practical, i.e. clinical impact of these approaches has been pretty small, and this is likely to continue to be true for foreseeable future To be fair, there are lots of non-technical and cultural issues behind this - not just failure to generalize.
It's asymptotic and it will never achieve 1:1 accuracy. The natural world doesn't have a measurable resolution, and this is apparent in written language, as we're seeing others detail in other comments, as well as it is in more relatable fields like sound. There will always be a difference between what your ears hear and 192kHz/24bit (and higher) digitized audio and/or video. That difference will always be a source of...mistakes.
Hey, it's just a few quantization errors. Nobody walking across a street or voting in an election has anything to worry about from those. /s
These are the fatal flaws, the hamartia if you will, in attempts to democratize technologies that are based on digitizing the real world and making decisions based on its interpretations of a quantized dataset. The opposite of the uncanny valley is you getting run over by a Waymo.
I wouldn't be too sure of that. People have shown time and time again that they are pretty bad at determining what a 'fact' is.
It gets far worse when they're shown more data (the issue we are dealing with right now), and these models see far more data than any one of us will see in our lifetimes.
Much of our ability to determine what a 'fact' is is either something like: (1) (if political) 'does it fit with my bias?' or (2) (if physics) 'is it physically possible?'. Since (1) is mostly dependent on what information you show yourself, and (2) depends on comparing to interaction with physical reality, the system has too much information for (1) and no physical space to run experiments in for (2). Much of our ability to determine what a 'fact' is is either something like: (1) (if political) 'does it fit with my bias?' or (2) (if physics) 'is it physically possible?'. Since (1) is mostly dependent on what information you show yourself, and (2) depends on comparing to interaction with physical reality, the system has too much information for (1) and no physical space to run experiments in for (2).
That's not a very interesting metric though. GPT-3 has access to all of wikipedia and has enough memory to store it all.
It's like saying a calculator is better at maths than a professional mathematician because it can multiply longer numbers.
Rightly? Delaying a life-saving measure when it's the #9 cause of death?
They should be deployed selectively as soon as they can make the most risk-prone situations safer than the cohort of involved human drivers. E.g. getting drunk people home. We don't have to wait until they surpass the average driver because the average driver is not necessarily the driver (heh) of deaths.
Doing anything else is leaving bodies on the table.
* Long term, the reputational damage to self driving cars might be significant if one is released that is only slightly better than a drunk driver. If this hinders uptake later, when self-driving cars are completely superior to normal humans, we've possibly produced a net negative.
* While a self driving car should be safer than a drunk driver, the best option would be to call a cab or have a designated driver. It seems morally fraught to provide a solution that is more dangerous than existing ones, even if the less dangerous solutions have less than 100% uptake.
* Related, there may be some people who will instead of viewing a self-driving car as some sort of emergency option to avoid putting another drunk driver on the road, see it as a more convenient option to inviting a designated driver.
* Issues around who is responsible when a self driving car which is known to be worse than a typical human is deployed, fails, and someone is hurt.
> Issues around who is responsible when a self driving car which is known to be worse than a typical human is deployed, fails, and someone is hurt.
Even if insurance rates would be initially(!) higher than for a regular taxi driver I would expect it to end up cheaper overall and save lives.
Also note that "worse-than-the-old-average" may still end up improving the average because the magnitude by which they're worse is lower. Also, there's nothing magical about "worse-than-average" because for all distributions (except the dirac distribution) there is always some fraction worse than the average, so we always have to expect some fraction of drivers to be worse than average and there's no fundamental reason why they must be human. In fact, if the fleet of all cars in the world consisted of very safe, very consistently behaving autonomous cars, except for a single vehicle driven by a very conservative human driver who only slowly circles in a fenced private community and never hits anyone then statistically the entire autonomous car fleet may be "worse than average" due to the single outlier.
I agree in the same way than 70 % of people have less 100 iq, we depend in specialist when I need to know if I have epilepsi I need a person/thing who work is be up to date, and have the less bias as possible and agregator models are quite usually miss in understand what is crital info, try to program only whit github copilot or translate a novel(they do probably better than I don't English native) but not nearly as translator, seems correct but it isn't
Being pedantic here, but isn't that more like ~50%, by definition? Or did I misunderstand how IQ works?
This is already well under way. It's called vector search[1]. Google, Bing, Facebook, Spotify, Amazon, etc etc already use this to power their search and recommender systems.
There are even a bunch of companies popping up (I work for one[2]) that let everyone else get in on the fun.
Check out this video with the creator of SBERT / SentenceTransformer explaining how vector search is used in combination with language models to power semantic search: https://youtu.be/7RF03_WQJpQ
[1] https://www.pinecone.io/learn/what-is-similarity-search/
A much harder problem is creating accurate vectors to begin with. Even the most advanced language models today create word/sentence embeddings that leave a lot to be desired. Not to mention this is slow and GPU intensive.
Creating an end-to-end solution for embedding/searching/ranking (of which vector search is just one component, the other one should be some kind of keyword based search to increase precision) is what would be very valuable to offer as a service.
I'm not sure what you mean by that. The definition is pretty clear cut as "any system that perceives its environment and takes actions that maximize its chance of achieving its goals". The media has started to get on the bandwagon of "it's not AI if it's not an AGI" which is just nonsense.
This is true, and in some regard, how it has always been.
https://en.wikipedia.org/wiki/AI_effect
Already, outside of obvious marketing, a lot of uses of the current wave of AI/deep learning have stopped becoming magical.
Reminds me of Feynman's anecdote about the bird and knowing the name of something https://www.youtube.com/watch?v=px_4TxC2mXU
No, memorizing trivia is not it.
Can GPT-3 prove theorems about Schwarzschild metric?
From the first (of three) linked papers in the article.
> Language modelling provides a step towards intelligent communication systems by harnessing large repositories of written human knowledge to better predict and understand the world. In this paper, we present an analysis of Transformer-based language model performance across a wide range of model scales — from models with tens of millions of parameters up to a 280 billion parameter model called Gopher. These models are evaluated on 152 diverse tasks, achieving state-of-the-art performance across the majority. Gains from scale are largest in areas such as reading comprehension, fact-checking, and the identification of toxic language, but logical and mathematical reasoning see less benefit. We provide a holistic analysis of the training dataset and model’s behaviour, covering the intersection of model scale with bias and toxicity. Finally we discuss the application of language models to AI safety and the mitigation of downstream harms.
The researchers seem to know what they're doing and not doing.
Microsoft is saying that more parameters is inherently better. I'm sure there's reasonable limits that they need to be competently implemented.
Left unsaid: it's much harder to learn shortcuts to fake performance in logic and reasoning and common arithmetic, than it is to do the same in "reading comprehension" and similar tasks for which there are no good benchmark datasets (most are basically multipl-choice tests that can be solved by a strong classifier without any need of anything like "comprehension") and for which there are no good metrics either (see BLEU score and friends).
As such there is no reason to expect them to be very different in term of efficiency, and it has been shown and well researched, that scaling the numbers of parameters directly correlates with improved model quality.
So as long as you are comparing GPT style models to other GPT style models then parameter count is definitly not a vanity metric.
This doesn't hold once you start comparing to e.g. mixture of experts model which were making the headline recently with trillion of parameters claims. In MoE models, parameter counts is pretty much a useless metric.
It evokes the feel of a technology that is impressive this year but is on the cusp of being overwhelmingly, cataclysmically, eclipsed very shortly by another, much more powerful, technology. In that previous case, Gopher and web of course.
I wonder whether this evocation was intended, as an aspect of the naming here in an AI context.
Attempting to do recreate the entirety of the physical world by describing it in text is an impressive feat, but a futile attempt.
It is prone to inference errors that the model itself is not aware of. There are less and less of such errors, but the very inability to tell that what you are outputting is gibberish is one of the biggest problems of these models. In other words, if the model could simply say 'I do not know' or 'I am not sure' for every occasion when it is not 100% sure in something (like a human can), this would drastically improve the usefulness.
[1] Spent a lot of time pondering on this topic https://arxiv.org/pdf/1906.01873.pdf
How did you reach this conclusion? Is there evidence or was it inferred from the model structure?
This is exactly what a language model does though, just at a different level of abstraction. It gives you a probability distribution over tokens at each step. That distribution can be narrow (low entropy, certain) or wide (high entropy, uncertain). The language output you see is just a sampling at some temperature from these distributions.
Though glancing at your paper I assume you are aware of this and I am missing the point you are making?
When a human does not know something it can tell that with 100% certainty.
That problem aside, language models like Gopher are in fact generative, so no such threshold is needed! You instead sample from the implicit distribution.
Equating model 'parameters' to interneuron connections in naïve at best (and a horrible measure in general).
All I'm trying to say is I find it crazy how dang big these models are getting.
I think you can add several orders of magnitude to that since nerve cells are more like microcontrollers (with memory, adaptation etc.) than simple nodes. I remember a scientific article that made a big impression on me: when a dragon fly sees a prey, only 8 neurons (connected to the eyes and the wings) are responsible for keeping it oriented toward the target.
But I did encounter this article a while ago here on HN.
Only two neurons are necessary to ride a bicycle.
A few orders of magnitude and an entire category away. Artificial "neurons" only have the name "neuron" in common with biological neurons. Consequently you can stack as many layers of artificial neurons on top of each other as you may want and you won't get anywhere near the abilities of the simplest systems of biological neurons.
For example, spiders have ~100 thousand neurons and there's no artificial neural networks that could shake a stick at a spider's cognitive abilities. Which are downright scary, btw.
Estimated number of neurons of spiders from wikipedia:
https://en.wikipedia.org/wiki/List_of_animals_by_number_of_n...
What actual value would an AI that produces answers similar to the average person have, though? Non-expert answers for interesting questions are pretty much meaningless -- the whole point of an advanced society is that we can avoid knowing anything about most things and focus on narrow expertise.
Very impressive, but not very useful to extract knowledge!
If they can do it, so can you.
If both the response to the corona virus, and the presidency of Trump have shown us ONE thing, it's that those who are higher up in the pyramid are NOT necessarily more clever or better informed than you are.
Not a good look to have an error on the landing page for your product.
They're transparently providing an example where their product doesn't work well. Find me another product, even an OSS project that does the same on their landing page.
I'm sure Google made lots of accusations if asked for comment, that seems SOP, but what actually triggered it? What documents did she leak?
Among OpenAI (and MS), Google and Nvidia, none of them have released their pretrained large language models.
Just like strong cryptography right? This is bullshit. Either it's plain dangerous and you shouldn't be doing it or it's morally fine and it should be open. Their final usecase is pushing ads or extorting data from unsuspecting people's interactions anyway so who are they kidding with their ethics..
I mean they'd be foolish giving it away: it's a big investment and a rare good, highly strategic. Let's not wrap that kind of action in some savior stance. They are just making money and it's ugly.
The final usecase is actually brainwashing. There could come a future where every single comment you read online is not a human but rather a multi-faceted AI tailored specifically for you. Your current worldview could be measured and the AI could steer you closer towards a "target" worldview.
The comments presented to you could also be moderately coordinated with comments presented to other humans in your real-life circles to establish "serendipity" where you and your friends/family/coworkers evolve similar ideas/worldviews together to reinforce your new beliefs. And even catalyze coordinated human efforts towards common goals.
Doing this could lead most people to believe almost anything. Theoretically this could also be coupled with VR chatrooms to enhance salience.
My best guess is that they want a reasonably coordinated release of their papers/blogs/tweets/YouTube-videos, and waiting for Arxiv's upload is unpredictable, so they simply host a copy themselves and the upload takes care of itself.
2. Success is not the only metric of worth. Nuclear power has "failed" despite being the only reliable solution to meeting the base load of a power grid vs fossil fuels.