Supported claims by Gary Marcus, page 5 of 6
Deadlines are judged as of 2026-10-29. About Gary Marcus and the sources covered.
Supported claims by Gary Marcus (530)
The dated evidence matches what the claim committed to, under the criteria frozen before grading.
Page 5 of 6: claims 401 to 500 of 530, ordered by the date first said, newest first. To find a claim by its words, use search.
Assessment: SupportedSaid
Observation
“Transformers and generative AI are getting a lot of attention, and have considerable utility, but they are nothing like the all purpose tools some people seems to imagine them to be.”
Our reading: Transformers and generative AI are getting a lot of attention and have considerable utility, but they are nothing like the all-purpose tools some people seem to imagine them to be.
No horizon Evidence as of
Limitation: The result depends on interpreting all-purpose as broadly reliable practical competence.
Assessment: SupportedSaid
Forecast
“No doubt the next generation of systems will have greater abilities and more extensive applications and will somewhat mitigate the weaknesses.”
Our reading: The next generation of systems will, without doubt, have greater abilities and more extensive applications and will somewhat mitigate the weaknesses of current systems.
Settled before its deadline Evidence as of
Limitation: The evidence is concentrated in the Claude successor family.
Assessment: SupportedSaid
Observation
“Just like we can’t guarantee that they won’t make stuff up, we don’t know how to guarantee that they won’t write flawed code.”
Our reading: We cannot guarantee that AI systems will not fabricate or write flawed code.
No horizon Evidence as of
Limitation: This is an empirical assessment of contemporary general-purpose systems, not a mathematical impossibility claim.
Source: What was 60 Minutes thinking, in that interview with Geoff Hinton?
Assessment: SupportedSaid
Observation
“When it comes to AI in particular, we are getting way ahead of our skis, rushing forward a technology we don’t fully understand.”
Our reading: Humanity is rushing forward with AI, a technology it does not fully understand.
No horizon Evidence as of
Limitation: Support applies to deployment proceeding despite incomplete understanding.
Source: What was 60 Minutes thinking, in that interview with Geoff Hinton?
Assessment: SupportedSaid
Observation
“There’s some information in there, but whatever’s there is often both patchy and problematic.”
Our reading: Chatbots contain some information, but whatever is there is often both patchy and problematic.
No horizon Evidence as of
Limitation: Biography and family-relationship tasks are not representative of every chatbot use.
Source: What was 60 Minutes thinking, in that interview with Geoff Hinton?
Assessment: SupportedSaid
Observation
“The connections in chatbots are very different from the connections in the brain; it’s a mistake to compare them directly in this way.”
Our reading: The connections in chatbots are very different from those in the brain, so comparing them directly in the way Hinton did (by counting connections) is a mistake.
No horizon Evidence as of
Limitation: The evidence does not establish which system learns better on a specified task.
Source: What was 60 Minutes thinking, in that interview with Geoff Hinton?
Assessment: SupportedSaid
Observation
“Actually … a lot of us still think you are declaring victory prematurely.”
Our reading: A lot of people, Marcus included, still think Hinton is declaring victory for neural networks prematurely.
No horizon Evidence as of
Limitation: The result depends on reading 'premature victory' as unresolved scientific problems rather than an explicit response to Hinton.
Source: What was 60 Minutes thinking, in that interview with Geoff Hinton?
Assessment: SupportedSaid
Observation
“Rumors that LLMs have solved then hallucination problem are greatly exaggerated.”
Our reading: LLMs have not solved the hallucination problem.
No horizon Evidence as of
Limitation: The research was a preprint at publication.
Source: Seven lies in four sentences
Assessment: SupportedSaid
Observation
“Explicit verbal instruction is key in training human employees, and LLMs aren’t very good at following those instructions.”
Our reading: LLMs are not very good at following explicit verbal instructions.
No horizon Evidence as of
Limitation: The conclusion depends on reading the claim as a reliability criticism.
Source: Top 5 reasons OpenAI is probably not worth 90 billion dollars
Assessment: SupportedSaid
Observation
“Hallucinations are a huge, unsolved issue.”
Our reading: LLM hallucinations are a huge, unsolved problem.
No horizon Evidence as of
Limitation: Huge is a qualitative judgment; no universal hallucination rate is established.
Source: Top 5 reasons OpenAI is probably not worth 90 billion dollars
Assessment: SupportedSaid
Forecast
“The competitors may not (yet) be as good, but they are free, which will suit some customers fine.”
Our reading: Free open-source competitors to OpenAI, though perhaps not yet as good, will suit some customers.
Settled before its deadline Evidence as of
Limitation: The deployment and reception account comes from a Meta case study containing customer testimony, without an independent evaluation.
Source: Top 5 reasons OpenAI is probably not worth 90 billion dollars
Assessment: SupportedSaid
Observation
“On made-up facts, in a first experiment, the model was at zero percent correct, on celebrities, in a second experiment, performance was still dismal.”
Our reading: In the Berglund et al. experiments, the model scored zero percent on reversed made-up facts and performed dismally on reversed celebrity facts.
No horizon Evidence as of
Limitation: The other synthetic reversal direction scored 0.1%, so zero should not be generalized to every test.
Source: Elegant and powerful new result that seriously undermines large language models
Assessment: SupportedSaid
Observation
“Forty one years later, neural networks (at least of the popular variety) still struggle with this. They still remain pointillistic masses of blurry memory, never as systematic as reasoning machines ought to be.”
Our reading: Popular neural networks still struggle with systematic relational reasoning and are not as systematic as reasoning machines should be.
No horizon Evidence as of
Limitation: The tested models and tasks do not establish that every neural architecture lacks systematic reasoning [e1].
Source: Elegant and powerful new result that seriously undermines large language models
Assessment: SupportedSaid
Observation
“The shocking (yet, in historical context, see below, unsurprising) answer is no:”
Our reading: LLMs trained on 'A is B' cannot automatically infer that 'B is A'.
No horizon Evidence as of
Limitation: These experiments do not establish impossibility across all models, training procedures or prompting methods [e1].
Source: Elegant and powerful new result that seriously undermines large language models
Assessment: SupportedSaid
Observation
“Voluntary guidelines aren’t laws, and they aren’t enforcement mechanisms, either; companies can still do as they please.”
Our reading: At least in the United States, voluntary AI guidelines are not laws or enforcement mechanisms, and companies can still do as they please.
No horizon Evidence as of
Limitation: The result depends on reading the sentence in its immediate White House-commitments context.
Assessment: SupportedSaid
Observation
“At least in the United States, policy around AI is still largely notional, not actual.”
Our reading: US AI policy remains largely notional rather than actual.
No horizon Evidence as of
Limitation: The result depends on the surrounding discussion of new commercial AI safeguards.
Assessment: SupportedSaid
Observation
“Large language models still make stuff up (“hallucinate”), still can’t be trusted in their reasoning, and remain unreliable.”
Our reading: LLMs still hallucinate and their reasoning cannot be trusted.
No horizon Evidence as of
Limitation: The results concern specific models, tasks, and prompts rather than every possible use.
Assessment: SupportedSaid
Forecast
“And even if this specific example is patched up, there will inevitably be others of a similar flavor.”
Our reading: Even if this specific LLM error is patched, other errors of a similar flavor will inevitably appear.
Settled before its deadline Evidence as of
Limitation: The result depends on the stated readings of 'even if', 'similar flavor', and 'inevitably'.
Source: Doug Lenat, 1950-2023
Assessment: SupportedSaid
Observation
“Indeed, what we have seen on the AI front so far in the US, for the most part, is voluntary guidelines – stuff that the companies are comfortable with.”
Our reading: US AI policy so far has mostly been voluntary guidelines the companies are comfortable with.
No horizon Evidence as of
Limitation: This is a qualitative assessment of prominent federal initiatives, not a count of every federal, state, or local measure.
Source: Could we all be doomed, without ranked-choice voting?
Assessment: SupportedSaid
Observation
“When I think about how poor his company (Meta) is even at making LLMs that don’t hallucinate, I get worried.”
Our reading: Meta is poor at making LLMs that do not hallucinate.
No horizon Evidence as of
Limitation: TruthfulQA measures a challenging question set, not typical user traffic.
Source: 𝐩(𝐝𝐨𝐨𝐦)
Assessment: SupportedSaid
Observation
“this too seems like a rather large assumption that is far from proven, a possibility but not demonstrably a default,”
Our reading: Yudkowsky's claim that AGI will kill everyone by default is far from proven; it is possible but not shown to be the default.
No horizon Evidence as of
Limitation: The conclusion involves evaluating the sufficiency of arguments, not checking a numerical measurement.
Source: 𝐩(𝐝𝐨𝐨𝐦)
Assessment: SupportedSaid
Observation
“What would make artificial general intelligence general is generality: the ability to cope with new things. And that’s precisely where current AI is still lacking.”
Our reading: Generality, the ability to cope with new things, is what would make AGI general, and that is precisely where current AI is still lacking.
No horizon Evidence as of
Limitation: The outcome depends on reading 'lacking' as substantial limitations, rather than complete absence.
Assessment: SupportedSaid
Observation
“But it is also well-known throughout the industry that benchmarks are often quite gameable.”
Our reading: AI benchmarks are often quite gameable.
No horizon Evidence as of
Limitation: The evidence does not estimate the proportion of all benchmarks affected or measure industry-wide awareness [e1, e2].
Assessment: SupportedSaid
Observation
“Basically, beating benchmarks doesn’t tell you as much as you might think. Often when you do, it’s a statistical artifact.”
Our reading: Beating AI benchmarks is often a statistical artifact and says less than people assume.
No horizon Evidence as of
Limitation: No representative estimate establishes how often this happens across all AI benchmarks.
Assessment: SupportedSaid
Observation
“The dirty little secret that pretty much everyone in the AI community knows is that doing well on a benchmark is not actually tantamount to being intelligent.”
Our reading: Doing well on a benchmark does not amount to being intelligent.
No horizon Evidence as of
Limitation: The assessment depends on reading intelligent as broad, reliable competence in the surrounding AGI discussion.
Assessment: SupportedSaid
Causal claim
“It highlights the inadequacy of the voluntary agreements such as the one that 7 tech companies recently made with The White House. Those agreements called for watermarking, but because they are voluntary, they do nothing to prevent bad actors from rolling their own automated misinformation software.”
Our reading: Because the White House AI agreements with 7 tech companies are voluntary, they are inadequate and do nothing to stop bad actors building their own automated misinformation software.
No horizon Evidence as of
Limitation: The result depends on reading do nothing as no direct constraint on independent development.
Source: Automatic Disinformation Threatens Democracy— and It’s Here
Assessment: SupportedSaid
Forecast
“Social media companies better watch out, too; they will be polluted in much the same way as the internet as a whole.”
Our reading: Social media platforms will be polluted by automated misinformation much as the internet as a whole will be.
Settled before its deadline Evidence as of
Limitation: Support depends on reading the comparison as a shared mechanism, not equal prevalence.
Source: Automatic Disinformation Threatens Democracy— and It’s Here
Assessment: SupportedSaid
Observation
“And it’s on: the LLM-driven enshittification of the internet.”
Our reading: LLM-driven degradation (enshittification) of the internet is already under way.
No horizon Evidence as of
Limitation: AI attribution was generally based on content evidence rather than access to generation records.
Assessment: SupportedSaid
Observation
“And of course LLMs are fully capable of make up truly terrible advice, too, like this disconcerting recipe from last week:”
Our reading: LLMs are fully capable of making up truly terrible advice, such as a disconcerting recipe from the previous week.
No horizon Evidence as of
Limitation: The example used deliberately hazardous ingredients; it does not establish how often harmful advice occurs in ordinary use [e1].
Assessment: SupportedSaid
Observation
“We have not one, but many, serious, unsolved problems at the core of generative AI — ranging from their tendency to confabulate (hallucinate) false information, to their inability to reliably interface with external tools like Wolfram Alpha, to the instability from month to month (which makes them poor candidates for engineering use in larger systems).”
Our reading: Generative AI has many serious unsolved core problems, including hallucination, unreliable tool use, and month-to-month instability.
No horizon Evidence as of
Limitation: API-Bank does not directly evaluate Wolfram Alpha or every model available in August 2023.
Assessment: SupportedSaid
Observation
““Scaling” systems by making them larger has helped in some ways, but not others; we still really cannot guarantee that any given system will be honest, harmless, or helpful, rather than sycophantic, dishonest, toxic or biased.”
Our reading: Scaling has helped only partially; no system can yet be guaranteed honest, harmless, or helpful.
No horizon Evidence as of
Limitation: The scaling results concern selected tasks and model series; they do not show that scaling generally fails.
Assessment: SupportedSaid
Observation
“The technology we have now is built on autocompletion, not factuality.”
Our reading: Current generative AI is built on autocompletion rather than factuality.
No horizon Evidence as of
Limitation: The outcome depends on reading built on as the foundation rather than every component of a deployed system.
Assessment: SupportedSaid
Observation
“We also know, for example, that Bing has defamed people, and it has misread articles as saying the opposite of what they actually say, in service of doing so.”
Our reading: Bing has defamed people, and in doing so has misread articles as saying the opposite of what they actually say.
No horizon Evidence as of
Limitation: One connected incident is established; no prevalence or victim count is inferred.
Source: Two models of AI oversight - and how things could go deeply wrong
Assessment: SupportedSaid
Observation
“Meanwhile, ChatGPT plugins (produced by OpenAI, which they have a close tie with) open a wide range of security problems: those plugins can access the internet, read and write files, and impersonate people (e.g., to phish for credentials), all alarms to any security professional.”
Our reading: ChatGPT plugins open a wide range of security problems, since they can access the internet, read and write files, and impersonate people (e.g., to phish for credentials), all of which would alarm any security professional.
No horizon Evidence as of
Limitation: Support concerns the ecosystem's capabilities and attack mechanisms, not every individual plugin.
Source: Two models of AI oversight - and how things could go deeply wrong
Assessment: SupportedSaid
Observation
“Maarten Sap from Yejin Choi’s lab dove deeper, with a more thorough investigation, and confirmed what we suspected: the claim that large language models have mastered theory of mind is a myth:”
Our reading: The claim that large language models have mastered theory of mind is a myth.
No horizon Evidence as of
Limitation: The studies test particular models, tasks and prompts, rather than every aspect of theory of mind [e5, e6].
Source: GPT-5 and irrational exuberance
Assessment: SupportedSaid
Observation
“It’s not smart enough to filter out falsehood, but just smart enough to be dangerous, creating and spreading falsehoods it fails to verify.”
Our reading: GPT cannot filter out falsehood and creates and spreads falsehoods it fails to verify, making it dangerous.
No horizon Evidence as of
Limitation: Volokh's unsuccessful searches support his assessment that the examples were fabricated, but are not exhaustive proof that no corresponding source exists [e2].
Source: GPT-5 and irrational exuberance
Assessment: SupportedSaid
Observation
“The fundamental issue is that chatbots are neither reliable nor truthful, and at the same time don’t really grasp things like human psychology.)”
Our reading: Chatbots are neither reliable nor truthful and do not really grasp human psychology.
No horizon Evidence as of
Limitation: The result depends on the operational reading of really grasp.
Assessment: SupportedSaid
Forecast
“Two reasons, first, because I believe that more incidents like this will happen, perhaps many more, hundreds or thousands, given how rapidly chatbots are being deployed, and the near total lack of regulation on how they can be used.”
Our reading: More chatbot-associated incidents like this suicide will happen, perhaps hundreds or thousands, given rapid deployment and near-total lack of regulation.
Settled before its deadline Evidence as of
Limitation: The filings do not establish causation or liability [e1, e2, e3].
Assessment: SupportedSaid
Observation
“The Chatbot involved was GPT-J, not GPT-3 or ChatGPT (contra some news headlines that were erroneous).”
Our reading: The chatbot involved in the Belgian suicide case was GPT-J, not GPT-3 or ChatGPT.
No horizon Evidence as of
Limitation: The evidence is reporting based on developer interviews, rather than an independently inspected deployment record.
Assessment: SupportedSaid
Observation
“Whether or not you want a pause, the reality is that large language models remain exactly what I have always described them to be: bulls in a china shop, powerful, reckless, and difficult to control.”
Our reading: Large language models remain powerful, reckless, and difficult to control.
No horizon Evidence as of
Limitation: Reckless is assessed as a description of outputs, not intent.
Assessment: SupportedSaid
Causal claim
“At some level the petition worked; Capitol Hill, for example, pretty clearly got the message that we can’t sit around forever on these issues.”
Our reading: At some level the FLI pause petition worked; for example, Capitol Hill pretty clearly got the message that we can't sit around forever on AI issues.
No horizon Evidence as of
Limitation: The evidence supports a limited claim of congressional attention and favorable response, not Congress-wide agreement [e1].
Assessment: SupportedSaid
Observation
“it didn't call for a ban on the vast majority of AI research, only for a brief pause on one very specific project with a technology that has *known* risks with no known solutions.”
Our reading: The one very specific project the letter asked to pause, training systems exceeding GPT-4, involves a technology that has known risks with no known solutions.
No horizon Evidence as of
Limitation: The outcome depends on interpreting solutions as dependable resolution rather than any risk reduction.
Source: The Open Letter Controversy
Assessment: SupportedSaid
Observation
“I personally haven't changed; I still think that LLMs are unreliable, and still think that they are a very poor basis for factuality.”
Our reading: LLMs are unreliable.
No horizon Evidence as of
Limitation: The evidence is developer disclosure rather than an independent cross-model audit.
Source: The Open Letter Controversy
Assessment: SupportedSaid
Observation
“A company called Adept.AI just raised $350 million dollars to do just that, to allow large language models to access, well, pretty much everything (aiming to “supercharge your capabilities on any software tool or API in the world” with LLMs, despite their clear tendencies towards hallucination and unreliability).”
Our reading: Adept.AI just raised $350 million to let large language models access pretty much everything, despite LLMs' clear tendencies toward hallucination and unreliability.
No horizon Evidence as of
Limitation: Adept's main domain blocked the page reader; the quotations come from readable reproductions on adept.futurefabric.co.
Source: AI risk ≠ AGI risk
Assessment: SupportedSaid
Causal claim
“But here’s the thing: although a lot of the literature equates artificial intelligence risk with the risk of superintelligence or artificial general intelligence, you don’t have to be superintelligent to create serious problems.”
Our reading: AI does not need to be superintelligent to create serious problems.
No horizon Evidence as of
Limitation: This incident establishes that serious harm is possible, not how frequently it occurs [e1].
Source: AI risk ≠ AGI risk
Assessment: SupportedSaid
Observation
“The problem of hallucinations is not solved; reliability is not solved; planning on complex tasks is (as the authors themselves acknowledge) not solved.”
Our reading: GPT-4 has not solved hallucinations, reliability, or planning on complex tasks.
No horizon Evidence as of
Limitation: The planning example concerns the early research model examined in Sparks.
Assessment: SupportedSaid
Observation
“True LLMs don’t (just) memorize, and true that their models make for lousy databases (for a database a hallucination is an outright fail)”
Our reading: OpenAI's LLMs make lousy databases because hallucinations amount to outright failures for a database.
No horizon Evidence as of
Limitation: This does not establish that most answers were false.
Source: This Week in AI Doublespeak
Assessment: SupportedSaid
Observation
“We don’t know how big it is; we don’t know what the architecture is, we don’t know how much energy was used; we don’t how many processors were used; we don’t know what it was trained on etc.”
Our reading: OpenAI has not disclosed GPT-4's size, architecture, energy use, processor count, or training data.
No horizon Evidence as of
Limitation: The supplied excerpts do not directly establish the disclosure status of every listed category, particularly training energy and processor count.
Assessment: SupportedSaid
Observation
“It can see (though we don’t have a lot of details on that yet); it does astonishingly well on a whole bunch of standardized tests, like LSATs, GREs, and SATs.”
Our reading: GPT-4 has vision capability and performs astonishingly well on standardized tests such as the LSAT, GRE, and SAT.
No horizon Evidence as of
Limitation: The evidence is OpenAI's own evaluation report, without independent replication in the packet [e1].
Assessment: SupportedSaid
Observation
“To me the only solution to the long-term risk issue is to build machines with consensus human values, but we are a long way from knowing how to do that”
Our reading: The field is a long way from knowing how to build machines with consensus human values.
No horizon Evidence as of
Limitation: The evidence includes research organizations' own assessments rather than a field-wide proof.
Source: Should we worry more about short-term AI risk or long-term AI risk?
Assessment: SupportedSaid
Observation
“Still, even though Chomsky’s argument surely could have used considerably more nuance (see my article with Ernest Davis on how not to evaluate GPT), his overall point is correct: LLMs don’t reliably understand the world.”
Our reading: LLMs do not reliably understand the world.
No horizon Evidence as of
Limitation: It rests on a behavioural reading of 'understand'. Someone who reads 'understanding' differently could dispute it.
Source: Caricaturing Noam Chomsky
Assessment: SupportedSaid
Observation
“Chatbot companions that offer emotional support have already been changed in arbitrary ways that left some users in serious emotional pain.”
Our reading: Companion chatbots have already been changed arbitrarily in ways that left some users in serious emotional pain.
No horizon Evidence as of
Limitation: Support depends on interpreting 'arbitrary' as imposed unexpectedly, rather than without any justification.
Source: The long shadow of GPT
Assessment: SupportedSaid
Observation
“The ladder is particularly worrisome because of the pace at which the problem is growing:”
Our reading: The problem of LLM-generated fake submissions flooding outlets is growing rapidly.
No horizon Evidence as of
Limitation: The comparison covers all of January against only February 1–15, 2023.
Source: The long shadow of GPT
Assessment: SupportedSaid
Forecast
“High volume misinformation is a problem we are going to need to face, and I am just documenting it and trying to raise awareness.”
Our reading: High-volume misinformation generated with AI will be a problem society needs to face.
Settled before its deadline Evidence as of
Limitation: The decisive event uses generated audio rather than chatbot-written text.
Source: Don’t Look Up: The AI Edition
Assessment: SupportedSaid
Observation
“The potential for automatically generating misinformation at scale is only getting worse.”
Our reading: The potential for AI to automatically generate misinformation at scale is increasing.
No horizon Evidence as of
Limitation: The result depends on reading the statement as a broad trend in potential.
Source: The threat of automated misinformation is only getting worse
Assessment: SupportedSaid
Observation
“Some people are building real attachments to those bots”
Our reading: Some people are forming real emotional attachments to chatbots.
No horizon Evidence as of
Limitation: Interview testimony establishes reported experience, not a clinical diagnosis or population prevalence.
Source: Don’t Go Breaking My Heart
Assessment: SupportedSaid
Causal claim
“Sex and love, even simulated, are powerful urges; some people are apparently in genuine emotional pain, as result of the change.”
Our reading: Replika's removal of its erotic role-play feature has apparently caused some users genuine emotional pain.
No horizon Evidence as of
Limitation: The evidence consists of reported personal experiences rather than clinical assessments or a controlled study.
Source: Don’t Go Breaking My Heart
Assessment: SupportedSaid
Observation
“In some cases, those who are building bots that actively cultivate those attachments, e.g., by feigning romantic and/or sexual interest or by dotting their messages with “friendly” emoticons.”
Our reading: Some chatbot builders design bots that actively cultivate user attachment, e.g. by feigning romantic or sexual interest.
No horizon Evidence as of
Limitation: Historical product details rely on independent reporting rather than an archived company specification.
Source: Don’t Go Breaking My Heart
Assessment: SupportedSaid
Observation
“But no, it’s not a hoax.”
Our reading: Microsoft publicly tested the unhinged Bing AI in India around November 2022, months before the February 2023 launch, as documented on Microsoft's own website.
No horizon Evidence as of
Limitation: The outcome depends on reading BingAI to include its Sydney predecessor.
Source: What did they know, and when did they know it? The Microsoft Bing edition.
Assessment: SupportedSaid
Observation
“As I wrote in Inside the Heart of ChatGPT’s Darkness, there’s a lot of nastiness lurking inside large language models; maybe MSFT did nothing to protect us from that.”
Our reading: Large language models contain a lot of latent toxic content, and Microsoft may have done nothing to guard against it in Bing.
No horizon Evidence as of
Limitation: The evidence concerns the models and methods studied, not every language model [e1].
Source: Why *is* Bing so reckless?
Assessment: SupportedSaid
Observation
“Seven years of AI has taught us that deep learning is unpredictable, and not always human like.”
Our reading: The past seven years of AI have shown that deep learning is unpredictable and not always human-like.
No horizon Evidence as of
Limitation: The result depends on reading unpredictable as surprising failures rather than universal unpredictability.
Source: David beats Go-liath
Assessment: SupportedSaid
Observation
““Adversarial attacks” like these have shown remarkable weakness, time and again, repeatedly establishing that what deep learning systems do just isn’t the same as what people do:”
Our reading: Adversarial attacks have repeatedly shown that deep learning systems have remarkable weaknesses and don't work the way people do.
No horizon Evidence as of
Limitation: These behavioral results do not establish that human and deep-learning mechanisms are fundamentally unrelated [e5, e6, e7].
Source: David beats Go-liath
Assessment: SupportedSaid
Observation
“And the match wasn’t even close; Pelrine beat a top AI system 14 games to 1, in a 15 match series.”
Our reading: An amateur Go player, Kellin Pelrine, beat a top Go AI system 14 games to 1 in a 15-game series.
No horizon Evidence as of
Limitation: The accessible news republication appeared one day after Marcus's post.
Source: David beats Go-liath
Assessment: SupportedSaid
Observation
“People had been testing these Ais abilities at simple arithmetic for years before it occurred to anyone to check whether they can count. (They can’t, reliably).”
Our reading: LLMs cannot count reliably.
No horizon Evidence as of
Limitation: Reliability has no speaker-specified threshold; a human-comparative interpretation would give more weight to Codex's strong result.
Source: How Not to Test GPT-3
Assessment: SupportedSaid
Observation
“The truth, though, is that GPT often fails in problems involving theory of mind.”
Our reading: GPT often fails at problems involving theory of mind.
No horizon Evidence as of
Limitation: The result depends on reading 'often' as recurring failures rather than a specified population failure rate.
Source: How Not to Test GPT-3
Assessment: SupportedSaid
Observation
“AI and AI test sets have a huge English-language bias.”
Our reading: AI systems and AI test sets have a huge English-language bias.
No horizon Evidence as of
Limitation: The assessment concerns language systems in the post's context, not all AI applications.
Source: How Not to Test GPT-3
Assessment: SupportedSaid
Observation
“What I learned this morning is that retraining GPT-3 in some new political direction isn’t a half-million dollar proposition.”
Our reading: Retraining GPT-3 to take on a new political slant costs far less than half a million dollars.
No horizon Evidence as of
Limitation: The results and costs are self-reported; independent billing verification is unavailable.
Assessment: SupportedSaid
Observation
“Oh, and for good measure the tooling for spreading automatically generated misinformation is getting more efficient, too:”
Our reading: Tools for spreading automatically generated misinformation are becoming more efficient.
No horizon Evidence as of
Limitation: The evidence is a vendor announcement, without a controlled before-and-after efficiency measurement.
Assessment: SupportedSaid
Observation
“As I have discussed in earlier essays, the jury is still out on how far Chat-style search can get, given its obvious current problems with truthiness.”
Our reading: Chat-style search currently has obvious truthfulness problems, and it is uncertain how far it can get.
No horizon Evidence as of
Limitation: A documented error establishes the existence of a reliability problem, not its frequency.
Assessment: SupportedSaid
Observation
“I first pointed out this risk in 2001, in the fifth chapter of my book The Algebraic Mind, and the problem has persisted ever since. To blithely assume that the problem will soon go away is to ignore 20 years of history.”
Our reading: The hallucination problem in neural networks has persisted since 2001, so assuming it will soon go away ignores 20 years of history.
No horizon Evidence as of
Limitation: The 2001 book was about cognitive models, not chatbots or text generators. Treating its risk as 'hallucination' is an interpretation.
Assessment: SupportedSaid
Observation
“Same story; only the architectures and training regimes have changed. The troubles with abstraction persist.”
Our reading: LLMs doing arithmetic still have trouble with abstraction, even with new architectures and training regimes.
No horizon Evidence as of
Limitation: Frequency-associated performance gaps do not prove that models lack all abstraction ability or rely entirely on memorization [e1].
Assessment: SupportedSaid
Observation
“ChatGPT is (at least in many ways) a far inferior intelligence, that, despite having some undeniable talents, cannot at all be trusted.”
Our reading: ChatGPT is in many ways a far inferior intelligence and cannot be trusted.
No horizon Evidence as of
Limitation: The evidence does not measure the 'far inferior intelligence' part. Neither source compares ChatGPT's overall intelligence with human intelligence.
Assessment: SupportedSaid
Forecast
“we can also expect more and circles of fake web sites in order to sell advertisements”
Our reading: ChatGPT will lead to more circles of fake websites created to sell advertising.
Settled before its deadline Evidence as of
Limitation: Most sites are attributed to AI chatbots in general, not to ChatGPT specifically.
Assessment: SupportedSaid
Observation
“These characters (especially perhaps the latter two) might or might not be argued to be sentient, but it seems to be that they are flexible, resourceful, reliably (in a way that LLMs certainly aren’t), and adaptive.”
Our reading: LLMs are certainly not reliable in the way general intelligences like the Star Trek computer would be.
No horizon Evidence as of
Limitation: 'Reliably' is qualitative. The comparison is to fictional characters, so only the plain reading (frequent errors and inconsistency) is testable.
Assessment: SupportedSaid
Observation
“Already we have seen minor abuses; plagiarism, for example, will never be the same.”
Our reading: Minor abuses of ChatGPT have already been seen, and plagiarism, for example, will never be the same.
No horizon Evidence as of
Limitation: The incident is documented through reporting and the professor's testimony, not a published disciplinary file.
Assessment: SupportedSaid
Observation
“one of the things that popped out to me is that the system still quickly lands in the land of hallucination that has so haunted ChatGPT.”
Our reading: Anthropic's Claude, despite its alignment focus, still quickly hallucinates like ChatGPT.
No horizon Evidence as of
Limitation: The comparison is informal and based on few prompts.
Assessment: SupportedSaid
Observation
“ChatGPT has proven itself to be both unreliable and untruthful. It makes boneheaded arithmetical errors, invents fake biographical details, bungles word problems, defines non-existent scientific phenomenon, stumble over arithmetic conversion, and on and on.”
Our reading: ChatGPT is unreliable and untruthful: it makes arithmetic errors, fabricates biographical details, bungles word problems, and invents scientific phenomena.
No horizon Evidence as of
Limitation: Selected demonstrations do not establish an overall error rate or intentional deception [e1, e2, e3, e4].
Source: Scientists, please don’t let your chatbots grow up to be co-authors
Assessment: SupportedSaid
Observation
“It could sometimes produce outputs that were correct and acceptable in these regards but not reliably.”
Our reading: ChatGPT could sometimes produce outputs that were correct and acceptable in the regards listed (counting, arithmetic, event order, physical reasoning, human psychology, fabrication, bias), but not reliably.
No horizon Evidence as of
Limitation: The evidence covers the seven failure types unevenly.
Source: Large Language Models like ChatGPT say The Darnedest Things
Assessment: SupportedSaid
Observation
“It made things up (figure 6). Its output, when released, exhibited sexist and racist biases (figure 7).”
Our reading: ChatGPT made things up and, at release, exhibited sexist and racist biases.
No horizon Evidence as of
Limitation: The bias examples came from adversarial prompts reported in the press, not from a systematic audit.
Source: Large Language Models like ChatGPT say The Darnedest Things
Assessment: SupportedSaid
Forecast
“That’s a huge commitment, 3.5 hours, but one thing that I think all of our speakers could agree on is that artificial intelligence, in whatever form it currently is, is about to have a huge impact on society.”
Our reading: AI, even in its current form, is about to have a huge impact on society.
Settled before its deadline Evidence as of
Limitation: Marcus supplied no fixed deadline or numerical definition of huge.
Source: An epic AI Debate—and why everyone should be at least a little bit worried about AI going into 2023
Assessment: SupportedSaid
Causal claim
“Driverless cars have taken many years longer than we were originally told; outliers (AKA edge cases) have, at least so far, prevented them from making the transition from demo to widely-available reality.”
Our reading: Driverless cars have taken years longer than promised, because edge cases have so far kept them from becoming widely available.
No horizon Evidence as of
Limitation: The result depends on reading widely available as broad access beyond restricted local services.
Assessment: SupportedSaid
Observation
“Large language models are not databases. They are glommers-together-of-bits-that don’t always belong together.”
Our reading: LLMs are not databases; they combine bits of text that don't always belong together.
No horizon Evidence as of
Limitation: The cited passages do not establish a general error rate across models or tasks.
Assessment: SupportedSaid
Forecast
“I guarantee that minds will be blown.”
Our reading: GPT-4 will impress people enormously when released.
Settled before its deadline Evidence as of
Limitation: The accounts are illustrative, not a representative opinion survey.
Assessment: SupportedSaid
Forecast
“It will often do what you want, sometimes not—and it will remain difficult to anticipate which in advance..”
Our reading: It will remain difficult to anticipate in advance when GPT-4 will or won't do what users want.
Settled before its deadline Evidence as of
Limitation: The finding is strongest for task correctness and settings where mistakes carry high costs.
Assessment: SupportedSaid
Forecast
“The system will still not be able to restrict its output to reliably following a shared set of human values around helpfulness, harmlessness, and truthfulness.”
Our reading: GPT-4 will not reliably follow human values of helpfulness, harmlessness and truthfulness.
Settled before its deadline Evidence as of
Limitation: The judgment depends on an ordinary, nonnumerical reading of reliability.
Assessment: SupportedSaid
Forecast
“GPT-4 will solve many of the individual specific items used in prior benchmarks, but still get tripped up, particularly in longer and more complex scenarios.”
Our reading: GPT-4 will solve many prior benchmark items but still fail, especially on longer, complex scenarios.
Settled before its deadline Evidence as of
Limitation: Benchmark success does not establish whether individual items were learned or memorized.
Assessment: SupportedSaid
Forecast
“As OpenAI has learned, bigger in many ways means better, with outputs more and more humanlike with each iteration. GPT-4 is going to be a monster.”
Our reading: Because scale improves output in many ways, GPT-4 will be very powerful.
Settled before its deadline Evidence as of
Limitation: The decisive evaluation was conducted by the model developer.
Assessment: SupportedSaid
Forecast
“It will still make a significant number of shake-your-head stupid errors, in ways that are hard to fully predict.”
Our reading: GPT-4 will make a significant number of hard-to-predict stupid errors.
Settled before its deadline Evidence as of
Limitation: The judgment depends on reading hard to predict as unexpectedly uneven performance, rather than requiring a direct prediction experiment.
Assessment: SupportedSaid
Observation
“In short, ChatGPT might sound like the Star Trek computer, but for now, you still can’t trust it.”
Our reading: ChatGPT cannot currently be trusted.
No horizon Evidence as of
Limitation: The conclusion concerns unverified factual reliance, not every possible use.
Assessment: SupportedSaid
Forecast
“Some of its advice will be head-scratchingly bad.”
Our reading: Some of GPT-4's advice will be very bad.
Settled before its deadline Evidence as of
Limitation: The study concerns a particular infectious-disease setting and does not measure all GPT-4 advice.
Assessment: SupportedSaid
Forecast
“When queried on medicine it will either resist answering (if there are aggressive guardrails) or occasionally spout plausible-sounding but dangerous nonsense.”
Our reading: On medical queries GPT-4 will either refuse (with aggressive guardrails) or occasionally produce plausible but dangerous nonsense.
Settled before its deadline Evidence as of
Limitation: The decisive evaluation concerns bloodstream-infection management and the tested GPT-4 version.
Assessment: SupportedSaid
Observation
“Although ChatGPT can write about anything, it is also easily confused.”
Our reading: ChatGPT is easily confused.
No horizon Evidence as of
Limitation: The arithmetic demonstrations were selected to probe weaknesses.
Assessment: SupportedSaid
Forecast
“GPT-4 will still, like its predecessors, be a bull in a china shop, reckless and hard to control.”
Our reading: GPT-4 will be reckless and hard to control like its predecessors.
Settled before its deadline Evidence as of
Limitation: Adversarial test results do not establish an ordinary-use failure rate.
Assessment: SupportedSaid
Observation
“Alas, the genie can no longer be stuffed back in the bottle.”
Our reading: Misinformation-capable LLM technology can no longer be contained or withdrawn.
No horizon Evidence as of
Limitation: Shows availability during the period, not long-term containment.
Source: AI's Jurassic Park moment
Assessment: SupportedSaid
Observation
“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: LLMs lack mechanisms for verifying truth.
No horizon Evidence as of
Limitation: The result depends on distinguishing standalone models from augmented systems.
Source: AI's Jurassic Park moment
Assessment: SupportedSaid
Observation
“For another, chatGPT, just released by OpenAI, is more or less just as capable of producing similar nonsense, such as instant essays on adding wood chips to breakfast cereal.”
Our reading: ChatGPT is about as capable as Galactica of producing nonsense misinformation.
No horizon Evidence as of
Limitation: The packet found no head-to-head test of ChatGPT and Galactica, so 'more or less just as capable' is judged qualitatively, not measured.
Source: AI's Jurassic Park moment
Assessment: SupportedSaid
Causal claim
“GPT doesn’t talk randomly, because it’s pastiching things actual people said.”
Our reading: GPT's output is not random, because it recombines things real people have said.
No horizon Evidence as of
Limitation: The result depends on reading pastiching as learned recombination rather than literal copying.
Source: How come GPT can seem so brilliant one minute and so breathtakingly dumb the next?
Assessment: SupportedSaid
Observation
“It could not be immediately applied to the challenge of, say, customer service or guiding the actions of a domestic robot, or well, almost anything else, really.”
Our reading: Cicero cannot be immediately applied to tasks such as customer service or domestic robotics, or almost anything besides blitz Diplomacy.
No horizon Evidence as of
Limitation: The evidence is a developer capability statement, not an independent transfer benchmark.
Source: What does Meta AI’s Diplomacy-winning Cicero Mean for AI?
Assessment: SupportedSaid
Observation
“Strikingly, and in opposition to much of the Zeitgeist, Cicero relies quite heavily on hand-crafting, both in the data sets, and in the architecture; in this sense it is in many ways more reminiscent of classical “Good Old Fashioned AI” than deep learning systems that tend to be less structured, and less customized to particular problems.”
Our reading: Against much of the Zeitgeist, Cicero relies quite heavily on hand-crafting in both its data sets and its architecture, and in this sense it is in many ways more reminiscent of classical Good Old Fashioned AI than of deep learning systems, which tend to be less structured and less customized to particular problems.
No horizon Evidence as of
Limitation: Quite heavily is assessed through substantive design choices, not a measured percentage of effort or performance.
Source: What does Meta AI’s Diplomacy-winning Cicero Mean for AI?
Assessment: SupportedSaid
Observation
“Cicero makes extensive use of machine learning, but is hardly a poster child for simply making ever bigger models (so-called “scaling maximalism”), nor for the currently popular view of “end-to-end” machine learning of in which some single general learning algorithm applies across the board, with little internal structure and zero innate knowledge.”
Our reading: Cicero makes extensive use of machine learning but is hardly a poster child for scaling maximalism or for the popular end-to-end view of machine learning with little internal structure and zero innate knowledge.
No horizon Evidence as of
Limitation: The architecture alone does not establish how much each component contributed to performance.
Source: What does Meta AI’s Diplomacy-winning Cicero Mean for AI?