Claims that can’t be checked by Gary Marcus, page 23 of 24
Deadlines are judged as of 2026-10-29. About Gary Marcus and the sources covered.
Claims that can’t be checked by Gary Marcus (2353)
Public evidence cannot settle these claims, so they are recorded and attributed, not scored. Most were screened out: both AI providers read the claim on its own and found it a value judgment, rhetoric, hedged, too vague to test, or about something no public evidence could show. The rest were found to be recommendations or too open to test in a full review.
Page 23 of 24: claims 2201 to 2300 of 2353, ordered by the date first said, newest first. To find a claim by its words, use search.
Assessment: Can’t be checkedSaid
Recommendation or value judgment
“In our view, it is past time to have something like this, given the inherent unpredictability and wide range of the systems.”
Our reading: In the authors' view, it is past time to have a community-facing corpus of ChatGPT and other LLM errors, given the systems' inherent unpredictability and wide range.
No horizon
Qualifier: in our view
Source: Large Language Models like ChatGPT say The Darnedest Things
Assessment: Can’t be checkedSaid
Recommendation or value judgment
“George, like Chomsky and myself, called for a greater emphasis on the understanding of human intelligence.”
Our reading: AI research should place greater emphasis on understanding human intelligence.
No horizon
Source: An epic AI Debate—and why everyone should be at least a little bit worried about AI going into 2023
Assessment: Can’t be checkedSaid
Forecast
“I, Gary Marcus, worried about whether contemporary approaches to AI would ever give solutions to four key aspects of thought that we ought to expect from any intelligent machine: reasoning, abstraction, compositionality, and factuality.”
Our reading: Contemporary approaches to AI may never solve reasoning, abstraction, compositionality, and factuality, which any intelligent machine should have.
No horizon
Source: An epic AI Debate—and why everyone should be at least a little bit worried about AI going into 2023
Assessment: Can’t be checkedSaid
Forecast
“Kai-Fu Lee, often far more optimistic than I am, worried about whether recent advances in AI-generated content were a net positive for society, or whether they might lead to an avalanche of misinformation [something I too am deeply worried about] and target advertising that might be detrimental.”
Our reading: Recent advances in AI-generated content may lead to an avalanche of misinformation.
No horizon
Source: An epic AI Debate—and why everyone should be at least a little bit worried about AI going into 2023
Assessment: Can’t be checkedSaid
Rhetoric
“If you want to know what a large cross section of the field’s leaders are thinking, right now, there’s no better place to start, a time capsule of 2022 as we enter into the remaining wild AI years of the 2020s.”
Our reading: The remaining years of the 2020s will be wild for AI.
No horizon
Source: An epic AI Debate—and why everyone should be at least a little bit worried about AI going into 2023
Assessment: Can’t be checkedSaid
Recommendation or value judgment
“David Ferrucci, CEO of Elemental Cognition, and director of IBM’s successful Watson Jeopardy effort, worried that current systems were “ultimately unsatisfied”, and like me called for hybrid approaches that combine neural networks with reasoning and structured representations.”
Our reading: AI should pursue hybrid approaches combining neural networks with reasoning and structured representations.
No horizon
Source: An epic AI Debate—and why everyone should be at least a little bit worried about AI going into 2023
Assessment: Can’t be checkedSaid
Causal claim
“The trouble is, those guardrails don’t come for free; the system is still shallow; ChatGPT doesn’t really know what it is it is guarding against; it mostly seems instead to be just looking for keywords, which is why it is prone to generating nonsense like this:”
Our reading: ChatGPT's guardrails don't really understand what they guard against and appear to rely mostly on keywords, which makes ChatGPT prone to producing nonsense.
No horizon
Assessment: Can’t be checkedSaid
Recommendation or value judgment
“There’s another problem, too, arguably even bigger concern in the context of search engines: hallucination.”
Our reading: For search engines, hallucination is a further problem for LLM-based systems and arguably an even bigger concern than the shallowness, or discomprehension, he described just before.
No horizon
Scope: in the context of search engines
Assessment: Can’t be checkedSaid
Recommendation or value judgment
“But until those bits all fit together in a way we can trust, I prefer to borrow a memorable phrase from Ariana Grande: Thank U, Next.”
Our reading: Until the pieces of LLM-based search fit together in a way that can be trusted, he personally prefers to pass on it ('Thank U, Next').
No horizon
Qualifier: until those bits all fit together in a way we can trust
Assessment: Can’t be checkedSaid
Forecast
“In the quest for a newfangled search I suspect we are once again in for a rough ride.”
Our reading: Developing LLM-based search will probably be a difficult, drawn-out process, as with driverless cars.
No horizon
Assessment: Can’t be checkedSaid
Recommendation or value judgment
“When it comes to LLM-driven search, caveat emptor!”
Our reading: Users should be cautious when relying on LLM-driven search.
No horizon
Assessment: Can’t be checkedSaid
Recommendation or value judgment
“It is insidious the way that truth and falsity are so thoroughly and authoritatively mixed together. I for one am not ready for our post-truth information overlords.”
Our reading: The way truth and falsity are thoroughly and authoritatively mixed together in LLM-driven search answers like the ones he examined (such as the Neeva bio) is insidious, and he personally is not ready for 'post-truth information overlords'.
No horizon
Scope: LLM-driven search outputs like the examples he examined, such as the Neeva bio
Assessment: Can’t be checkedSaid
Observation
“I like to call this sort of nonsense discomprehension: mindless answers that show that the systems has no clue what it is that is talking about.”
Our reading: ChatGPT's mindless answers ('discomprehension') show it has no clue what it is talking about.
No horizon
Assessment: Can’t be checkedSaid
Forecast
“I suspect that what we will see is a familiar pattern: immense initial buzz, followed by a more careful scientific inspection, followed by a recognition that many problems remain.”
Our reading: GPT-4 will likely generate immense buzz followed by scientific scrutiny and recognition that many problems remain.
No horizon
Assessment: Can’t be checkedSaid
Conditional forecast
“When AGI (artificial intelligence) comes, large language models like GPT-4 may be seen in hindsight as part of the eventual solution, but only as part of the solution.”
Our reading: When AGI comes, large language models like GPT-4 may be seen in hindsight as part of the eventual solution, but only as part of it.
No horizon
Only if: when AGI comes
Assessment: Can’t be checkedSaid
Recommendation or value judgment
“If all seven predictions prove correct, I hope that the field will finally realize that it is time to move on.”
Our reading: Marcus hopes that, if all seven of his predictions prove correct, the field will finally realize that it is time to move on.
No horizon
Only if: if all seven predictions prove correct
Assessment: Can’t be checkedSaid
Forecast
“And it’s clear that that is exactly what large language models make possible. They are aiming to create a world in which we are unable to know what we can trust; with these new tools, they might succeed.”
Our reading: LLMs make greatly escalated propaganda volume possible, and propagandists might succeed in creating a world where people cannot know what to trust.
No horizon
Source: AI's Jurassic Park moment
Assessment: Can’t be checkedSaid
Rhetoric
“Executives at Meta and OpenAI are as enthusiastic about their tools as the proprietors of Jurassic Park were about theirs.”
Our reading: Executives at Meta and OpenAI are as enthusiastic about their tools as the proprietors of Jurassic Park were about theirs.
No horizon
Source: AI's Jurassic Park moment
Assessment: Can’t be checkedSaid
Observation
“It is no exaggeration to say that systems like these pose a real and imminent threat to the fabric of society.”
Our reading: Generative AI systems like chatGPT pose a real and imminent threat to the fabric of society.
No horizon
Source: AI's Jurassic Park moment
Assessment: Can’t be checkedSaid
Recommendation or value judgment
“Large language models lack mechanisms for verifying truth; we need to find new ways to integrate them with the tools of classical AI, such as databases, webs of knowledge, and reasoning.”
Our reading: We need to integrate LLMs with classical AI tools such as databases, knowledge graphs, and reasoning to fight misinformation.
No horizon
Source: AI's Jurassic Park moment
Assessment: Can’t be checkedSaid
Forecast
“Nation-states and other bad actors that deliberately produce propaganda, are highly unlikely to voluntarily put down their new arms. Instead, they are likely to use large language models as a new class of automatic weapons, in their war on truth, attacking social media and crafting fake web sites at a volume we have never seen before.”
Our reading: Propaganda-producing nation-states and bad actors are likely to use LLMs to attack social media and create fake websites at unprecedented volume.
No horizon
Source: AI's Jurassic Park moment
Assessment: Can’t be checkedSaid
Recommendation or value judgment
“In time, though it would not be a popular decision, we may have to begin to treat misinformation as we do libel, making it actionable if it is created with sufficient malice and sufficient volume.”
Our reading: Countries may need to make misinformation legally actionable like libel when created with sufficient malice and volume.
No horizon
Qualifier: if created with sufficient malice and volume
Source: AI's Jurassic Park moment
Assessment: Can’t be checkedSaid
Recommendation or value judgment
“First, every social media company and search engine should support and extend StackOverflow’s ban; automatically-generated content that is misleading, should not be welcome, and the regular posting of it should be grounds for a user’s removal.”
Our reading: Every social media company and search engine should support and extend Stack Overflow's ban, misleading automatically-generated content should not be welcome, and regularly posting it should be grounds for a user's removal.
No horizon
Scope: automatically-generated content that is misleading
Source: AI's Jurassic Park moment
Assessment: Can’t be checkedSaid
Recommendation or value judgment
“User accounts must be more strenuously validated, and new systems like Harvard and Mozilla’s human-ID.org that allow for anonymous, bot-resistant authentication need to become mandatory; they are no longer a luxury we can afford to wait on.”
Our reading: User accounts must be more strongly validated and anonymous bot-resistant authentication should become mandatory.
No horizon
Source: AI's Jurassic Park moment
Assessment: Can’t be checkedSaid
Conditional forecast
“As one of the most central resources that the world’s programmers rely on, the consequences for software quality and developer productivity could be immense.”
Our reading: If Stack Overflow is flooded with worthless ChatGPT-generated code examples and dies, then, because it is one of the most central resources the world's programmers rely on, the consequences for software quality and developer productivity could be immense.
No horizon
Only if: if Stack Overflow is flooded with worthless generated code examples and dies
Source: AI's Jurassic Park moment
Assessment: Can’t be checkedSaid
Observation
“GPT-3 has no idea how the world works (and on this Bender and I would agree);”
Our reading: GPT-3 does not understand how the world works.
No horizon
Source: How come GPT can seem so brilliant one minute and so breathtakingly dumb the next?
Assessment: Can’t be checkedSaid
Observation
“But it is mimic that knows not whereof it speaks.”
Our reading: GPT-3 is a mimic that does not understand what it is saying.
No horizon
Source: How come GPT can seem so brilliant one minute and so breathtakingly dumb the next?
Assessment: Can’t be checkedSaid
Recommendation or value judgment
“It was just named a Best Tech Article of 2022, and I believe that is still incredibly relevant, even with all the advances of the last 8 months.”
Our reading: Marcus believes his essay 'Deep Learning is Hitting a Wall' is still incredibly relevant, even with all the advances of the last 8 months.
No horizon
Qualifier: I believe
Source: How come GPT can seem so brilliant one minute and so breathtakingly dumb the next?
Assessment: Can’t be checkedSaid
Recommendation or value judgment
“The results themselves are, without question, genuinely impressive.”
Our reading: Cicero's Diplomacy results are, without question, genuinely impressive.
No horizon
Qualifier: without question
Source: What does Meta AI’s Diplomacy-winning Cicero Mean for AI?
Assessment: Can’t be checkedSaid
Causal claim
“(The contest was a “blitz” version of the game, limited to five minutes a move, which favors computers, since the blitz version is presumably less dialogue- and negotiation-intensive than the full version of the game; in a longer version of the game, humans may still dominate.)”
Our reading: The five-minute-a-move blitz version of Diplomacy favors computers, presumably because it is less dialogue- and negotiation-intensive than the full game, and in a longer version humans may still dominate.
No horizon
Qualifier: presumably
Source: What does Meta AI’s Diplomacy-winning Cicero Mean for AI?
Assessment: Can’t be checkedSaid
Conditional forecast
“Our best guess is that if you played Diplomacy with alternative rules, the system would like to have retrained almost from scratch,”
Our reading: If Diplomacy were played with alternative rules, Cicero would probably need to be retrained almost from scratch.
No horizon
Only if: if Diplomacy were played with alternative rules
Source: What does Meta AI’s Diplomacy-winning Cicero Mean for AI?
Assessment: Can’t be checkedSaid
Forecast
“Cicero may change that calculus.”
Our reading: Cicero may change the field's view of machine learning as a universal solvent.
No horizon
Source: What does Meta AI’s Diplomacy-winning Cicero Mean for AI?
Assessment: Can’t be checkedSaid
Observation
“Cicero is in many ways a marvel; it has achieved by far the deepest and most extensive integration of language and action in a dynamic world of any AI system built to date.”
Our reading: Cicero is in many ways a marvel and has achieved by far the deepest and most extensive integration of language and action in a dynamic world of any AI system built to date.
No horizon Evidence as of
Limitation: This does not dispute that Cicero represented a substantial technical achievement.
Source: What does Meta AI’s Diplomacy-winning Cicero Mean for AI?
Assessment: Can’t be checkedSaid
Conditional forecast
“as things stand now, the prospects for generalization seem to us to be somewhat limited; the style of work might well be useful in other problems, but it may be that not much of the specifics of the architecture would survive if the system were applied to other problems”
Our reading: Saying he cannot predict this with any great confidence, Marcus judges that, as things stand now, Cicero's prospects for generalization seem somewhat limited: its style of work might well be useful in other problems, but it may be that not much of its specific architecture would survive if the system were applied to other problems.
No horizon
Only if: if the system were applied to other problems
Source: What does Meta AI’s Diplomacy-winning Cicero Mean for AI?
Assessment: Can’t be checkedSaid
Forecast
“This one may well go on for a long time, advancing as technology itself advances.”
Our reading: The argument over whether AI systems are sentient may well go on for a long time, evolving as the technology advances.
No horizon
Qualifier: advancing as technology advances
Source: Sentience and AI: A Dialog between Gary Marcus and Blake Lemoine
Assessment: Can’t be checkedSaid
Observation
“Consciousness is probably like that; we probably can’t really answer the questions we are striving for, at least not yet; we have to satisfy ourselves with trying to understand the arguments clearly.”
Our reading: Questions about consciousness probably cannot yet really be answered, given current science and technology, so we must settle for understanding the arguments clearly.
No horizon
Qualifier: at least not yet
Source: Sentience and AI: A Dialog between Gary Marcus and Blake Lemoine
Assessment: Can’t be checkedSaid
Forecast
“they are not (and may never be) great tech for reliably inferring user intent from what users say.”
Our reading: LLMs may never become great technology for reliably inferring user intent.
No horizon
Assessment: Can’t be checkedSaid
Causal claim
“LLMs are inherently unreliable. If Alexa were to make frequent errors, people would stop using it.”
Our reading: Marcus suspects that a contributing reason Alexa does not converse is that LLMs are inherently unreliable, and if Alexa made frequent errors, people would stop using it.
No horizon
Only if: if Alexa were to make frequent errors
Assessment: Can’t be checkedSaid
Causal claim
“Any one of these scenarios of LLMs gone awry (bad advice, insults, lies etc) could hurt the Amazon brand, open up litigation, etc..”
Our reading: Any one of the scenarios of LLMs going awry (bad advice, insults, lies, etc.) could hurt the Amazon brand and open up litigation.
No horizon
Only if: any one of these scenarios of LLMs gone awry
Assessment: Can’t be checkedSaid
Observation
“Amazon probably *is* using large language models for product search.”
Our reading: Amazon is probably using large language models for product search.
No horizon
Assessment: Can’t be checkedSaid
Recommendation or value judgment
“No matter how much data you can collect, don’t count on deep learning.”
Our reading: Twitter should not count on deep learning to detect misinformation, regardless of how much data it collects.
No horizon
Qualifier: regardless of data quantity
Source: Bravo @ElonMusk for wanting Twitter to become the world’s “most accurate source of information”
Assessment: Can’t be checkedSaid
Recommendation or value judgment
“You are going to need AI. But not (just) the popular stuff.”
Our reading: Twitter will need AI to fight misinformation, but not just popular AI approaches.
No horizon
Source: Bravo @ElonMusk for wanting Twitter to become the world’s “most accurate source of information”
Assessment: Can’t be checkedSaid
Recommendation or value judgment
“You’re going to need something different.”
Our reading: Fighting misinformation on Twitter requires an AI approach different from deep learning/LLMs.
No horizon
Source: Bravo @ElonMusk for wanting Twitter to become the world’s “most accurate source of information”
Assessment: Can’t be checkedSaid
Conditional forecast
“The proof makes too many assumptions to be taken as gospel, but if it is even close to correct, there may soon be real trouble in Scaling City.”
Our reading: If the Merrill, Warstadt and Linzen proof is even close to correct, scaling maximalism may soon be in real trouble.
No horizon
Only if: if the Merrill/Warstadt/Linzen proof is even close to correct
Source: What “game over” for the latest paradigm in AI might look like
Assessment: Can’t be checkedSaid
Recommendation or value judgment
“If you get syntax and semantics but fail on pragmatics or commonsense reasoning, you don’t have AGI you can trust.”
Our reading: A system that gets syntax and semantics right but fails on pragmatics or commonsense reasoning is not AGI you can trust.
No horizon
Scope: system gets syntax and semantics but fails on pragmatics or commonsense reasoning
Source: What “game over” for the latest paradigm in AI might look like
Assessment: Can’t be checkedSaid
Recommendation or value judgment
“All of this points to the same conclusion: if we want to reach AGI we shouldn’t keep putting so many eggs in the scaling-über-alles basket.”
Our reading: If the goal is AGI, the AI field should stop putting so much of its effort into scaling alone.
No horizon
Qualifier: if we want to reach AGI
Source: What “game over” for the latest paradigm in AI might look like
Assessment: Can’t be checkedSaid
Forecast
“There might not be enough data in the world to make scaling maximalism work.”
Our reading: There may not be enough data in the world for scaling maximalism to work.
No horizon
Source: What “game over” for the latest paradigm in AI might look like
Assessment: Can’t be checkedSaid
Conditional forecast
“What hit me hard, as I was reading the paper, is that asymptotic 80% performance on even a single important task like this might spell game over for scaling.”
Our reading: If performance plateaus at about 80% on even one important task like pragmatics, that may mean game over for scaling.
No horizon
Only if: if performance plateaus at about 80% on a single important task
Source: What “game over” for the latest paradigm in AI might look like
Assessment: Can’t be checkedSaid
Forecast
“There might not be enough available compute in the world to make scaling maximalism feasible.”
Our reading: There may not be enough available compute in the world to make scaling maximalism feasible.
No horizon
Source: What “game over” for the latest paradigm in AI might look like
Assessment: Can’t be checkedSaid
Forecast
“Some important tasks might simply not scale.”
Our reading: Some important tasks may simply not improve with scale.
No horizon
Source: What “game over” for the latest paradigm in AI might look like
Assessment: Can’t be checkedSaid
Causal claim
“Knowing more about exactly what it is beyond size that makes our mind unique would tell us an immense amount, both about ourselves, and about how we might want to build our AI.”
Our reading: Knowing more about exactly what it is, beyond brain size, that makes the human mind unique would tell us an immense amount both about ourselves and about how we might want to build AI.
No horizon
Qualifier: knowing more about exactly what beyond size makes our mind unique
Assessment: Can’t be checkedSaid
Forecast
“Undoubtedly clever intelligences than ours are possible and may yet materialize.”
Our reading: Intelligences cleverer than human intelligence are possible and may eventually materialize.
No horizon
Assessment: Can’t be checkedSaid
Recommendation or value judgment
“MetaAI’s new text-to-movie software Make a Video is straight-up amazing.”
Our reading: Meta AI's new text-to-video software Make a Video is straight-up amazing.
No horizon
Assessment: Can’t be checkedSaid
Causal claim
“i disagree w this logic: building biggest model isn’t inherently a contribution to the innovations we need”
Our reading: Having the most data and compute and building the biggest model is not inherently a contribution to the innovations needed for AGI.
No horizon
Source: Sub-Optimal
Assessment: Can’t be checkedSaid
Conditional forecast
“With respect to raw computational power, Tesla might eventually outgun Boston Dynamics, depending on how things go, but I seriously doubt they could outgun Google if Google went all in on humanoid robotics.”
Our reading: Tesla might eventually exceed Boston Dynamics in compute, but very likely could not outgun Google if Google went all in on humanoid robotics.
No horizon
Only if: if Google went all in on humanoid robotics
Source: Sub-Optimal
Assessment: Can’t be checkedSaid
Observation
“(Though he clearly has not invested nearly enough so far.)”
Our reading: Musk clearly has not, so far, invested nearly enough to win the robotics race.
No horizon
Scope: so far
Source: Sub-Optimal
Assessment: Can’t be checkedSaid
Observation
“Meta, which has more data and compute than almost anyone, has finally woken up to that. Musk still hasn’t.”
Our reading: Meta has recognized that data and compute alone won't suffice, while Musk has not.
No horizon
Source: Sub-Optimal
Assessment: Can’t be checkedSaid
Conditional forecast
“With enough investment, Tesla might in principle catch up. If Musk really wants to win the robotics race, he has the resources to do so.”
Our reading: With enough investment, Tesla might in principle catch up to Boston Dynamics in humanoid robotics.
No horizon
Only if: with enough investment
Source: Sub-Optimal
Assessment: Can’t be checkedSaid
Observation
“The most important concern is that robotics is always harder than people think it is.”
Our reading: Robotics is always harder than people think it is.
No horizon
Source: Top 5 reasons Tesla’s Optimus announcement may turn out be a dud
Assessment: Can’t be checkedSaid
Conditional forecast
“Roboticists are likely to scoff unless they see real evidence of progress than on Friday.”
Our reading: Roboticists are likely to scoff at Tesla's Friday Optimus showing unless they see real evidence of progress.
No horizon
Only if: unless they see real evidence of progress
Source: Top 5 reasons Tesla’s Optimus announcement may turn out be a dud
Assessment: Can’t be checkedSaid
Forecast
“The fleets of driverless taxis he promised are likely still years or decades away.”
Our reading: The fleets of Tesla driverless taxis Musk promised are likely still years or decades away as of September 2022.
No horizon
Source: Top 5 reasons Tesla’s Optimus announcement may turn out be a dud
Assessment: Can’t be checkedSaid
Forecast
“None of this is to say that Tesla could not solve all of these problems eventually given enough time and money, but Friday may not be the big demo his fans are hoping for.”
Our reading: Tesla's Optimus demo on Friday, September 30, 2022 may not be the big demo Musk's fans are hoping for.
No horizon
Source: Top 5 reasons Tesla’s Optimus announcement may turn out be a dud
Assessment: Can’t be checkedSaid
Recommendation or value judgment
“In my view the Google demo is pretty problematic (despite being cool), particularly in terms of likely reliability, but no matter.”
Our reading: In Marcus's view, Google's demo hooking a large language model to a robot, though cool, is pretty problematic, particularly in terms of likely reliability.
No horizon
Qualifier: in my view
Source: Top 5 reasons Tesla’s Optimus announcement may turn out be a dud
Assessment: Can’t be checkedSaid
Recommendation or value judgment
“A large part of LeCun’s new manifesto is a well-motivated call for incorporating a “configurable predictive world model” into deep learning.”
Our reading: Deep learning should incorporate configurable predictive world models.
No horizon
Assessment: Can’t be checkedSaid
Observation
“I won’t accuse LeCun of plagiarism, because I think he probably reached these conclusions honestly, after recognizing the failures of current architectures.”
Our reading: Marcus thinks LeCun probably reached his conclusions honestly, after recognizing the failures of current architectures, and so he will not accuse LeCun of plagiarism.
No horizon
Qualifier: I think he probably
Assessment: Can’t be checkedSaid
Conditional forecast
“It may be boring to hear about the same challenges over and over, but it’s hard to see how we will get to trustworthy general intelligence until these problems are resolved, in ways that are robust, rather than merely statistically contingent.”
Our reading: It is hard to see how we will get to trustworthy general intelligence until the challenges he has repeatedly raised are resolved in ways that are robust rather than merely statistically contingent.
No horizon
Only if: until these problems are resolved in ways that are robust rather than merely statistically contingent
Assessment: Can’t be checkedSaid
Rhetoric
“Those who wish that AI criticism would go away are welcome to their views, but misrepresenting critics and their views (and obscuring the degree to which even mainstream researchers agree) won’t magically solve the problems AI has yet to solve.”
Our reading: Misrepresenting AI critics and their views, and obscuring the degree to which even mainstream researchers agree, won't magically solve the problems AI has yet to solve.
No horizon
Assessment: Can’t be checkedSaid
Recommendation or value judgment
“In assessing progress towards general intelligence, the critical question should be, how much do systems like Dall-E, Imagen, Midjourney, and Stable Diffusion really understand the world, such that they can reason on and act on that knowledge?”
Our reading: Progress towards general intelligence from image systems like DALL-E, Imagen, Midjourney and Stable Diffusion should be assessed by how much they understand the world so they can reason and act on it.
No horizon
Qualifier: in assessing progress towards general intelligence
Source: Form, function, and the giant gulf between drawing a picture and understanding the world
Assessment: Can’t be checkedSaid
Observation
“(Dare I make obvious guess? They probably tried and failed, and don’t want to admit it.)”
Our reading: Google probably tested Imagen on Winoground, got poor results, and doesn't want to admit it.
No horizon
Source: Did GoogleAI Just Snooker One of Silicon Valley’s Sharpest Minds?
Assessment: Can’t be checkedSaid
Observation
“For me, compositionality has always been the touchstone for AI; when I say AI has hit a wall, it’s compositionality, more than anything else, that I have been talking about:”
Our reading: AI has hit a wall, and that wall is mainly compositionality.
No horizon
Source: Did GoogleAI Just Snooker One of Silicon Valley’s Sharpest Minds?
Assessment: Can’t be checkedSaid
Recommendation or value judgment
“A will to believe in AI will never replace the need for careful science.”
Our reading: A will to believe in AI will never replace the need for careful science.
No horizon
Source: Did GoogleAI Just Snooker One of Silicon Valley’s Sharpest Minds?
Assessment: Can’t be checkedSaid
Recommendation or value judgment
“Having a Silicon Valley celebrity like Alexander weigh in for Google is great PR for Google, but not so great for the public understanding of AI.”
Our reading: Having a Silicon Valley celebrity like Alexander weigh in for Google is great PR for Google but not so great for the public understanding of AI.
No horizon
Source: Did GoogleAI Just Snooker One of Silicon Valley’s Sharpest Minds?
Assessment: Can’t be checkedSaid
Recommendation or value judgment
“In 2017, a team of Google researchers invented the Transformer, arguably the biggest advance in AI in the last 10 years.”
Our reading: The Transformer, invented by a team of Google researchers in 2017, is arguably the biggest advance in AI in the last 10 years.
No horizon
Qualifier: arguably
Source: Did GoogleAI Just Snooker One of Silicon Valley’s Sharpest Minds?
Assessment: Can’t be checkedSaid
Conditional forecast
“Without a much greater degree of interpretability and verifiability, its hard to see how we can get to that level of safety.”
Our reading: Without a much greater degree of interpretability and verifiability, it is hard to see how humanoid home robots can reach the level of safety needed to act reliably and safely on users' requests.
No horizon
Only if: without a much greater degree of interpretability and verifiability
Assessment: Can’t be checkedSaid
Conditional forecast
“Misunderstandings will arise; without really firm checks on these systems, some of those misunderstandings could lead to mayhem.”
Our reading: LLM-driven robots will misunderstand people, and without firm checks some of those misunderstandings could lead to mayhem.
No horizon
Only if: without really firm checks on these systems
Assessment: Can’t be checkedSaid
Conditional forecast
“It’s not just that large language models can counsel suicide, or sign off on genocide, or that they can be toxic and that they are incredibly (over)sensitive to the details of their training set - it’s that if you put them inside a robot, and they misunderstand you, or fail to fully appreciate the implications of your request, they could cause major damage.”
Our reading: If large language models are put inside robots and misunderstand a request or miss its implications, they could cause major damage.
No horizon
Only if: LLMs put inside a robot
Assessment: Can’t be checkedSaid
Causal claim
“And the more actions a robot might consider, the worse performance is likely to be.”
Our reading: The more actions a robot like PaLM-SayCan has to consider, the worse its performance is likely to be.
No horizon
Assessment: Can’t be checkedSaid
Rhetoric
“Putting aside robots for the moment, we already know that so-called large language models are like bulls in a china shop: awesome, powerful, and reckless.”
Our reading: Large language models are powerful but reckless, like bulls in a china shop.
No horizon
Assessment: Can’t be checkedSaid
Forecast
“First, the language technology that the new system relies on is well-known to be problematic and second it is likely to be even more problematic in the context of robots.”
Our reading: Language model technology is likely to be even more problematic when used in robots.
No horizon
Qualifier: in the context of robots
Assessment: Can’t be checkedSaid
Recommendation or value judgment
“In a general-purpose, humanoid domestic robot, 75% is nowhere near good enough.”
Our reading: In a general-purpose humanoid domestic robot, a 75% success rate is nowhere near good enough.
No horizon
Scope: general-purpose humanoid domestic robot
Assessment: Can’t be checkedSaid
Recommendation or value judgment
“The PaLM-SayCam experiment makes it pretty clear that the list of 21 problems is due for an update.”
Our reading: DeepMind's list of 21 ethical and social problems with LLMs should be updated to cover the risks of putting LLMs in robots.
No horizon
Assessment: Can’t be checkedSaid
Causal claim
“As I have argued throughout the essays in this Substack (eg. Three ideas from linguistics that everyone in AI should know and The New Science of Alt Intelligence) , that’s because large language models are superficial statistical mimics, rather than systems that traffic in rich cognitive models of the world around them.”
Our reading: The field has no compelling idea at present for solving the many alignment problems of large language models because those models are superficial statistical mimics rather than systems that use rich cognitive models of the world.
No horizon
Qualifier: at the current time
Assessment: Can’t be checkedSaid
Observation
“The reality is that the field has no compelling idea at the current time how to solve the many “alignment” problems that bedevil large language models.”
Our reading: The AI field currently has no compelling idea of how to solve the alignment problems of large language models.
No horizon
Qualifier: at the current time
Assessment: Can’t be checkedSaid
Observation
“But that precisely what Google’s new system is, a stitching-together of superficial and incorrigible language understanders with powerful and potentially dangerous humanoid robots.”
Our reading: Google's PaLM-SayCan stitches superficial, incorrigible language models onto powerful and potentially dangerous humanoid robots.
No horizon
Assessment: Can’t be checkedSaid
Recommendation or value judgment
“I am not surprised that this can (kinda sorta) be done, but I am not sure it should be done.”
Our reading: Marcus doubts that language-model-driven robots like PaLM-SayCan should be built, even though they can sort of be made to work.
No horizon
Assessment: Can’t be checkedSaid
Observation
“And it is worth noting that even if they did work, there would be, of course, no guarantee that they would work in a fashion that was similar to humans; AlphaGo plays a mean game of Go but that doesn’t mean it gives insight into how human players learn Go.)”
Our reading: Even if systems like GPT-3 did work, there would be no guarantee that they would work in a fashion similar to humans.
No horizon
Qualifier: even if they did work
Assessment: Can’t be checkedSaid
Recommendation or value judgment
“If we want to make progress, we need to start by recognizing the enormity of the problem.”
Our reading: To make progress on language acquisition (human or machine), researchers must first recognize how hard the problem is.
No horizon
Qualifier: if we want to make progress
Assessment: Can’t be checkedSaid
Recommendation or value judgment
“Moving beyond prediction into reference, particularly in a system that is inherently uninterpretable, is an immense leap that (a) cannot be taken for granted and (b) should be a focus of research.”
Our reading: Moving beyond prediction into reference, especially in a system that is inherently uninterpretable, is an immense leap that cannot be taken for granted and should be a focus of research.
No horizon
Qualifier: particularly in a system that is inherently uninterpretable
Source: Three ideas from linguistics that everyone in AI should know
Assessment: Can’t be checkedSaid
Observation
“Rather, it means identify the second of the balls that are blue, regardless of which position; systems like GPT-3 likely have trouble with this sort of thing, as Ernie Davis just reported to us in an email.”
Our reading: Systems like GPT-3 likely have trouble with compositional phrases such as 'the second blue ball'.
No horizon
Source: Three ideas from linguistics that everyone in AI should know
Assessment: Can’t be checkedSaid
Forecast
“Considerable efforts are being made towards making “cross-modal” models like Gato that take in visual input as well strings of words; we think this is well worth investigating but unlikely in itself to suffice.”
Our reading: Cross-modal models like Gato are worth investigating but are unlikely by themselves to solve the lack of reference.
No horizon
Qualifier: in itself
Source: Three ideas from linguistics that everyone in AI should know
Assessment: Can’t be checkedSaid
Forecast
“For that, we may yet need another revolution.”
Our reading: Getting to real language understanding may require another revolution in AI.
No horizon
Source: Three ideas from linguistics that everyone in AI should know
Assessment: Can’t be checkedSaid
Rhetoric
“Yet even more wild is the widespread belief that language models lacking the major foundational components of human language (e.g., bounded compositionality mixed with unbounded generativity) will have any hope of meaningfully interfacing with audio/visual models in a way that accurately reflects human conceptual structure.”
Our reading: The authors call it even more wild than the expected unification of modalities in foundation models that people widely believe language models lacking major foundational components of human language (e.g., bounded compositionality mixed with unbounded generativity) have any hope of meaningfully interfacing with audio/visual models in a way that accurately reflects human conceptual structure.
No horizon
Scope: language models lacking the major foundational components of human language
Source: Three ideas from linguistics that everyone in AI should know
Assessment: Can’t be checkedSaid
Causal claim
“A good language understanding (or language production) system almost certainly needs to have some large-scale learning as a vital component.”
Our reading: A good language understanding or language production system almost certainly needs some large-scale learning as a vital component.
No horizon Evidence as of
Limitation: The evidence does not establish that the proposed necessity is true or false.
Source: Three ideas from linguistics that everyone in AI should know
Assessment: Can’t be checkedSaid
Causal claim
“But if all a system is doing is predicting the next word in what someone says, its understanding is inevitably superficial.”
Our reading: A system that only predicts the next word will inevitably have only superficial understanding.
No horizon
Qualifier: if the system only predicts the next word
Source: Three ideas from linguistics that everyone in AI should know
Assessment: Can’t be checkedSaid
Observation
“In our view, large language models are basically sidestepping this entire set of issues, and to their peril.”
Our reading: Large language models sidestep the problems of reference and shared understanding, and this puts them at risk.
No horizon
Source: Three ideas from linguistics that everyone in AI should know
Assessment: Can’t be checkedSaid
Forecast
“Ultimately, the right approach to large-scale learning may not rest on machinery that is finely geared to predicting sequences of words, but rather on developing a new approach to machine learning that is fundamentally geared towards learning how hierarchically structured sets of words map onto meanings that are used in the service of updating cognitive models.”
Our reading: The right approach to large-scale learning may need a new kind of machine learning that maps hierarchically structured words onto meanings used to update cognitive models, rather than word-sequence prediction.
No horizon
Source: Three ideas from linguistics that everyone in AI should know
Assessment: Can’t be checkedSaid
Recommendation or value judgment
“A couple weeks ago I got so fed up with so many smart people spending so much time experimenting with GPT-3 and DALL-E and so forth (when I, personally, think that there are deeper problems to be solved), that I wrote a poem.”
Our reading: Marcus thinks smart people are spending too much time experimenting with GPT-3 and DALL-E when there are deeper problems in AI to solve.
No horizon
Source: Poetry for Our AGI Times
Assessment: Can’t be checkedSaid
Recommendation or value judgment
“For example, we lack data on whether drivers use such systems as directed, and need to enforce vigorous mechanisms to make sure drivers don’t tune out.”
Our reading: Vigorous mechanisms should be enforced to ensure drivers using Level 2 driver-assist don't tune out.
No horizon
Assessment: Can’t be checkedSaid
Conditional forecast
“Clearly, the number of incidents would go up, as the number of vehicles went up, and would also go up significantly as the systems were more regularly challenged with more difficult driving conditions (bad weather, more erratic pedestrians, etc), and likely jump radically if we took humans out of the loop altogether.”
Our reading: If the whole US fleet used current self-driving software, incidents would rise with more vehicles and harder conditions, and likely jump radically without humans in the loop.
No horizon
Only if: if the entire US fleet used such software
Assessment: Can’t be checkedSaid
Recommendation or value judgment
“Is that good, or bad? I don’t know it, but it’s enough to be concerning. Why? Because .75 incidents a day comes from a very limited trial of computer-assisted driving relative to the ambition of full-self-driving.”
Our reading: Although he does not know whether it is good or bad, Tesla's roughly 0.75 ADAS incidents per day is enough to be concerning because it comes from a very limited trial of computer-assisted driving relative to the ambition of full self-driving.
No horizon
Qualifier: he does not know whether the rate is good or bad