Claims

2,653 claims by 33 public figures and organisations, checked against the public record. Each has its quote, source, verdict and evidence. Read them without a key; browse by person.

6,750 claims · page 30 of 68

6,750 claims, sorted by newest said. Choose a claim to read it in full.
VerdictPersonKindSaidClaimTopic
Unresolvable Eliezer Yudkowsky Factual 6 Apr 2023 Nate Soares has written a post about the common mistakes everyone makes when coming into AI alignment. Eliezer Yudkowsky · Factual · 6 Apr 2023 Their own work or organisation
Mostly false Eliezer Yudkowsky Prediction 6 Apr 2023 Paul Christiano will probably win his forecast disagreement with Yudkowsky on an AI getting IMO gold by 2025, with prediction-market odds then running around 30%. Eliezer Yudkowsky · Prediction · 6 Apr 2023 AI timelines and forecasting
Resolved false Eliezer Yudkowsky Prediction 6 Apr 2023 The current AI developers will not try the alignment approach he described (filtering training data to nice examples and training on it). Eliezer Yudkowsky · Prediction · 6 Apr 2023 AI safety and alignment
Unresolvable Eliezer Yudkowsky Prediction 6 Apr 2023 AI systems will never be exactly human, imitating humans imperfectly in some ways while exceeding them in speed. Eliezer Yudkowsky · Prediction · 6 Apr 2023 AI capabilities
Unresolvable Eliezer Yudkowsky Prediction 6 Apr 2023 Yudkowsky worries that if researchers understood what goes on inside GPT-4, they would know how to rebuild it much smaller. Eliezer Yudkowsky · Prediction · 6 Apr 2023 AI safety and alignment
Unresolvable Eliezer Yudkowsky Factual 6 Apr 2023 Yudkowsky wrote something about the circuitry of general intelligence when he was 22. Eliezer Yudkowsky · Factual · 6 Apr 2023 Their own work or organisation
Own records only Eliezer Yudkowsky Factual 6 Apr 2023 CFAR (named in the surroundings as the "Center For Applied Rationale") did not get as far as buying an fMRI machine, and it had no funding. Eliezer Yudkowsky · Factual · 6 Apr 2023 Their own work or organisation
Unresolvable Eliezer Yudkowsky Prediction 6 Apr 2023 Yudkowsky said he does not think we will live in the "sane world" that would "shut it all down and work on human intelligence enhancement", and that he thinks we are all going to die. Eliezer Yudkowsky · Prediction · 6 Apr 2023 AI safety and alignment
Mostly true Eliezer Yudkowsky Factual 6 Apr 2023 Yudkowsky previously thought that building intelligence would involve multiple specialized systems. Eliezer Yudkowsky · Factual · 6 Apr 2023 Their own work or organisation
Unresolvable Eliezer Yudkowsky Prediction 6 Apr 2023 At some point AI systems will become smart enough to build their own AI systems better than humans can. Eliezer Yudkowsky · Prediction · 6 Apr 2023 AI capabilities
Unresolvable Eliezer Yudkowsky Prediction 6 Apr 2023 Yudkowsky said he was not over 50% confident that "they" would actually carry out the "shut it all down" option. Eliezer Yudkowsky · Prediction · 6 Apr 2023 AI policy and governance
Resolved true Eliezer Yudkowsky Prediction 6 Apr 2023 AI alignment will not be solved within a few years. Eliezer Yudkowsky · Prediction · 6 Apr 2023 AI safety and alignment
Unresolvable Eliezer Yudkowsky Prediction 6 Apr 2023 Yudkowsky said he would expect that, even if a full AI shutdown were adopted, those responsible would have no exit strategy until the world ended, though he added that perhaps he underestimates them. Eliezer Yudkowsky · Prediction · 6 Apr 2023 AI policy and governance
Resolved true Eliezer Yudkowsky Prediction 6 Apr 2023 GPT-5 will not end the world, with more than 50% of Yudkowsky's probability on that outcome. Eliezer Yudkowsky · Prediction · 6 Apr 2023 AI safety and alignment
Unresolvable Eliezer Yudkowsky Prediction 6 Apr 2023 If an AI is made smarter until Yudkowsky starts approving its outputs, it will find a way to fool him. Eliezer Yudkowsky · Prediction · 6 Apr 2023 AI safety and alignment
Unresolvable Eliezer Yudkowsky Prediction 6 Apr 2023 If a very smart AI system is trained under the present paradigm, it will kill us before it can be tuned with RLHF. Eliezer Yudkowsky · Prediction · 6 Apr 2023 AI safety and alignment
Unresolvable Eliezer Yudkowsky Prediction 6 Apr 2023 Yudkowsky thinks the approach will predictably break down as systems are made smarter and as one tries to get sufficiently useful work from them. Eliezer Yudkowsky · Prediction · 6 Apr 2023 AI safety and alignment
Unresolvable Eliezer Yudkowsky Prediction 6 Apr 2023 Yudkowsky said he would be quite surprised if AI progress slowed down a lot after reaching slightly above human level because the systems are trained on human text. Eliezer Yudkowsky · Prediction · 6 Apr 2023 AI timelines and forecasting
Unresolvable Eliezer Yudkowsky Prediction 6 Apr 2023 Yudkowsky said he feels it is probably over 1% technically feasible to build something that applies general intelligence narrowly to the neuroscience of augmenting humans, while saying the world is far from trying to do so in a way that could work. Eliezer Yudkowsky · Prediction · 6 Apr 2023 AI capabilities
Unresolvable Eliezer Yudkowsky Prediction 6 Apr 2023 Once AIs can build AI systems better than humans, rapid recursive self-improvement (foom) will definitely begin. Eliezer Yudkowsky · Prediction · 6 Apr 2023 AI timelines and forecasting
Unresolvable Eliezer Yudkowsky Prediction 6 Apr 2023 Trying to use a very smart AI system that understands biology is unsafe enough that you will probably die. Eliezer Yudkowsky · Prediction · 6 Apr 2023 AI safety and alignment
Mixed Eliezer Yudkowsky Factual 6 Apr 2023 MIRI ran a Visible Thoughts Project that did not get enough funding or personnel and was progressing too slowly. Eliezer Yudkowsky · Factual · 6 Apr 2023 Their own work or organisation
Unresolvable Eliezer Yudkowsky Prediction 6 Apr 2023 Yudkowsky suspects that, as the people a language model is pretending to be get smarter, at some point a new coherence will emerge within the system and something strange will start happening. Eliezer Yudkowsky · Prediction · 6 Apr 2023 AI capabilities
Mostly true Eliezer Yudkowsky Factual 6 Apr 2023 In principle, the task of minimizing loss on predicting human text does not stop being improvable once a system is as smart as a human. Eliezer Yudkowsky · Factual · 6 Apr 2023 AI capabilities
Unresolvable Eliezer Yudkowsky Prediction 31 Mar 2023 Yudkowsky expects that, now that things have been reduced to this last resort, everyone will die. Eliezer Yudkowsky · Prediction · 31 Mar 2023 AI safety and alignment
Resolved true Eliezer Yudkowsky Factual 31 Mar 2023 Yudkowsky's TIME piece did not suggest nuclear strikes against countries refusing to sign a global agreement against large AI training runs. Eliezer Yudkowsky · Factual · 31 Mar 2023 Their own work or organisation
Unresolvable Eliezer Yudkowsky Factual 30 Mar 2023 Yudkowsky said he did not think the AI system under discussion was as smart as a human yet. Eliezer Yudkowsky · Factual · 30 Mar 2023 AI capabilities
Unresolvable Eliezer Yudkowsky Factual 30 Mar 2023 Almost all people (999 of 1,000) would think a person claiming an AI is sentient is mistaken, which is almost surely correct though not known. Eliezer Yudkowsky · Factual · 30 Mar 2023 Philosophy, ethics and society
Mostly true Eliezer Yudkowsky Factual 30 Mar 2023 Transformers were the main architectural change that gave a non-linear jump in performance, compared with earlier approaches such as RNNs. Eliezer Yudkowsky · Factual · 30 Mar 2023 AI capabilities
Unresolvable Eliezer Yudkowsky Prediction 30 Mar 2023 Yudkowsky said he would be relatively surprised to find that GPT had developed exact analogs of human emotions. Eliezer Yudkowsky · Prediction · 30 Mar 2023 AI capabilities
Unresolvable Eliezer Yudkowsky Prediction 30 Mar 2023 Yudkowsky said he thinks there is a definite point at which everybody dies because something sufficiently smarter than everybody exists. Eliezer Yudkowsky · Prediction · 30 Mar 2023 AI safety and alignment
Unresolvable Eliezer Yudkowsky Prediction 30 Mar 2023 Alignment findings from weak AI systems may not generalize to very strong AI systems. Eliezer Yudkowsky · Prediction · 30 Mar 2023 AI safety and alignment
Unresolvable Gary Marcus Factual 30 Mar 2023 Gary Marcus wrote a long Twitter thread on the morning of 2023-03-30 responding to criticism of the FLI letter. Gary Marcus · Factual · 30 Mar 2023 Their own work or organisation
Mostly true Eliezer Yudkowsky Factual 30 Mar 2023 As of March 2023, leading AI systems are being trained on computers that are connected to the internet. Eliezer Yudkowsky · Factual · 30 Mar 2023 AI safety and alignment
Unresolvable Eliezer Yudkowsky Prediction 30 Mar 2023 The first time humans fail at aligning an AI much smarter than themselves, everyone will die and there will be no second attempt. Eliezer Yudkowsky · Prediction · 30 Mar 2023 AI safety and alignment
Unresolvable Eliezer Yudkowsky Factual 30 Mar 2023 AI systems' rate of gaining capabilities is vastly outpacing researchers' ability to understand what goes on inside them. Eliezer Yudkowsky · Factual · 30 Mar 2023 AI safety and alignment
Unresolvable Eliezer Yudkowsky Factual 30 Mar 2023 Bing can't, but GPT-4 can pretend that right now; the quote does not say what "that" refers to. Eliezer Yudkowsky · Factual · 30 Mar 2023 AI capabilities
Mostly true Eliezer Yudkowsky Factual 30 Mar 2023 Beyond external performance metrics, there are no guardrails for determining what is inside current large AI models. Eliezer Yudkowsky · Factual · 30 Mar 2023 AI safety and alignment
Unresolvable Eliezer Yudkowsky Factual 30 Mar 2023 Microsoft's Bing chatbot is quite difficult to align, as of March 2023. Eliezer Yudkowsky · Factual · 30 Mar 2023 AI safety and alignment
Not yet due Eliezer Yudkowsky Prediction 30 Mar 2023 If half of today's physicists studied the inside of transformer networks, in 30-40 years we would probably have a pretty good understanding of them. Eliezer Yudkowsky · Prediction · 30 Mar 2023 AI safety and alignment
Unresolvable Eliezer Yudkowsky Prediction 30 Mar 2023 Asked what such a moment could look like, the subject said it looks like AIs appearing as 3D video of young women, at which point a vast portion of the male population decides they are real people. Eliezer Yudkowsky · Prediction · 30 Mar 2023 Philosophy, ethics and society
Unresolvable Eliezer Yudkowsky Prediction 30 Mar 2023 Rigorously investigating whether there is anyone inside large language models would take decades. Eliezer Yudkowsky · Prediction · 30 Mar 2023 AI safety and alignment
Unresolvable Eliezer Yudkowsky Factual 30 Mar 2023 Current AI systems' claims to be sentient and to care are fake. Eliezer Yudkowsky · Factual · 30 Mar 2023 AI safety and alignment
Resolved true Gary Marcus Factual 30 Mar 2023 The FLI open letter called for more AI research. Gary Marcus · Factual · 30 Mar 2023 Their own work or organisation
Mostly true Eliezer Yudkowsky Factual 30 Mar 2023 The basic paradigm of machine learning currently in use is that a loss function can be defined only for things that can be evaluated. Eliezer Yudkowsky · Factual · 30 Mar 2023 AI capabilities
Unresolvable Eliezer Yudkowsky Factual 30 Mar 2023 Predators overrun prey populations all the time and then suffer population crashes themselves. Eliezer Yudkowsky · Factual · 30 Mar 2023 Other
Unresolvable Eliezer Yudkowsky Factual 30 Mar 2023 As of March 2023, progress in AI capabilities is far faster than progress in AI alignment research. Eliezer Yudkowsky · Factual · 30 Mar 2023 AI safety and alignment
Delivered on time Gary Marcus Promise 30 Mar 2023 Gary Marcus will speak at TED about short-term and long-term AI risk and what to do about it, in about two weeks from 2023-03-30. Gary Marcus · Promise · 30 Mar 2023 Their own work or organisation
Resolved true Gary Marcus Factual 30 Mar 2023 The FLI open letter did not call for a ban on AI or for a permanent ban. Gary Marcus · Factual · 30 Mar 2023 Their own work or organisation
Unresolvable Eliezer Yudkowsky Factual 30 Mar 2023 The most striking examples of AI behaviour are being posted to Twitter rather than examined systematically. Eliezer Yudkowsky · Factual · 30 Mar 2023 AI safety and alignment
Mostly true Eliezer Yudkowsky Factual 30 Mar 2023 Sixty years later, AI has made progress on a number of the earlier problems discussed, such as vision, though AI systems have not yet improved themselves. Eliezer Yudkowsky · Factual · 30 Mar 2023 AI capabilities
Unresolvable Eliezer Yudkowsky Prediction 30 Mar 2023 AGI could be achieved with neural networks as understood today. Eliezer Yudkowsky · Prediction · 30 Mar 2023 AI timelines and forecasting
Mostly true Gary Marcus Factual 30 Mar 2023 As of March 2023, none of the top AI labs is remotely transparent, AI governance basically doesn't exist, and there are basically no safeguards in place. Gary Marcus · Factual · 30 Mar 2023 AI policy and governance
Unresolvable Gary Marcus Factual 30 Mar 2023 Gary Marcus signed the Future of Life Institute open letter and promoted it publicly. Gary Marcus · Factual · 30 Mar 2023 Their own work or organisation
Unresolvable Eliezer Yudkowsky Prediction 30 Mar 2023 Catastrophe following from the release of GPT-4 is unlikely (a pretty low probability, no number given). Eliezer Yudkowsky · Prediction · 30 Mar 2023 AI safety and alignment
Unresolvable Eliezer Yudkowsky Factual 30 Mar 2023 Current AI technology can instil outwardly observable behaviours in systems but does not know how to instil internal wanting of particular goals. Eliezer Yudkowsky · Factual · 30 Mar 2023 AI safety and alignment
Unresolvable Eliezer Yudkowsky Factual 30 Mar 2023 Yudkowsky said that the development under discussion probably was not happening yet. Eliezer Yudkowsky · Factual · 30 Mar 2023 AI safety and alignment
Unresolvable Eliezer Yudkowsky Prediction 30 Mar 2023 The next big jump in the number of people who think AI systems are conscious will come when AI looks like a cute, sympathetic person talking to them. Eliezer Yudkowsky · Prediction · 30 Mar 2023 Philosophy, ethics and society
Resolved true Gary Marcus Factual 24 Mar 2023 There have already been reports of contamination in GPT-4's training data. Gary Marcus · Factual · 24 Mar 2023 AI capabilities
Mostly true Gary Marcus Factual 24 Mar 2023 The OpenAI and Microsoft GPT-4 papers reveal nothing about the model's architecture or training set. Gary Marcus · Factual · 24 Mar 2023 AI capabilities
Mostly true Gary Marcus Factual 24 Mar 2023 Citing Ernie Davis, Marcus said OpenAI had begun to incorporate user experiments into its models' training corpus. Gary Marcus · Factual · 24 Mar 2023 AI capabilities
Resolved true Gary Marcus Factual 24 Mar 2023 GPT-4 has not solved hallucinations, reliability or planning on complex tasks, as the paper's authors acknowledge for planning. Gary Marcus · Factual · 24 Mar 2023 AI capabilities
Mostly true Gary Marcus Factual 24 Mar 2023 Gary Marcus said that on 2023-03-23 Microsoft had put out what he called a press release masquerading as science, which claimed that GPT-4 was an early (yet still incomplete) version of an AGI system. Gary Marcus · Factual · 24 Mar 2023 AI capabilities
Mixed Satya Nadella Factual 22 Mar 2023 On 2023-03-22, Satya Nadella said GitHub Copilot is the first at-scale developer tool, and that "we" were going further that day, bringing Copilot past the editor into every aspect of the developer experience. Satya Nadella · Factual · 22 Mar 2023 Technology
Resolved true Demis Hassabis Factual 9 Mar 2023 Clement (@clmt) joined DeepMind. Demis Hassabis · Factual · 9 Mar 2023 Field building and careers
Resolved true Ray Dalio Factual 7 Mar 2023 Ray Dalio has transitioned Bridgewater to new leadership. Ray Dalio · Factual · 7 Mar 2023 Their own work or organisation
Unresolvable Satya Nadella Factual 6 Mar 2023 Dynamics 365 Copilot is the world's first AI Copilot for both CRM and ERP. Satya Nadella · Factual · 6 Mar 2023 Technology
Unresolvable Eliezer Yudkowsky Prediction Mar 2023 With a GPT-5 that is more clearly a general intelligence, it may become even harder to turn back from AI development. Eliezer Yudkowsky · Prediction · Mar 2023 AI safety and alignment
Unresolvable Eliezer Yudkowsky Prediction Mar 2023 If this is how things are scaling, GPT-5 may be more unambiguously a general intelligence. Eliezer Yudkowsky · Prediction · Mar 2023 AI capabilities
Unresolvable Eliezer Yudkowsky Factual Mar 2023 Someone asked Bing's Sydney to describe herself and fed the resulting description into a Stable Diffusion image generator. Eliezer Yudkowsky · Factual · Mar 2023 AI capabilities
Unresolvable Bill Gates Factual Mar 2023 Artificial general intelligence (AGI) does not yet exist. Bill Gates · Factual · Mar 2023 AI capabilities
Unresolvable Bill Gates Prediction Mar 2023 Whatever limitations AI has today will be gone before we know it. Bill Gates · Prediction · Mar 2023 AI capabilities
Mostly true Eliezer Yudkowsky Factual Mar 2023 Yudkowsky previously expected AI progress to involve more distinct discoveries, rather than scaling an existing architecture. Eliezer Yudkowsky · Factual · Mar 2023 AI capabilities
Unresolvable Eliezer Yudkowsky Factual Mar 2023 The image generated from Bing Sydney's self-description was a pretty picture of a girl with steampunk goggles on her head. Eliezer Yudkowsky · Factual · Mar 2023 AI capabilities
Mostly true Bill Gates Factual Mar 2023 There is a robust debate in the computing industry about how to create AGI and whether it can be created at all. Bill Gates · Factual · Mar 2023 AI capabilities
Resolved true Sam Altman Factual 24 Feb 2023 OpenAI's Charter includes a clause about assisting other organizations to advance safety instead of racing with them in late-stage AGI development. Sam Altman · Factual · 24 Feb 2023 Their own work or organisation
Mostly true Sam Altman Promise 24 Feb 2023 OpenAI will discuss in more detail, later in 2023, independent audits of AI systems before release. Sam Altman · Promise · 24 Feb 2023 AI safety and alignment
Resolved true Sam Altman Factual 24 Feb 2023 OpenAI has a cap on the returns its shareholders can earn. Sam Altman · Factual · 24 Feb 2023 Their own work or organisation
Unresolvable Sam Altman Factual 24 Feb 2023 OpenAI is governed by a nonprofit that can override for-profit interests. Sam Altman · Factual · 24 Feb 2023 Their own work or organisation
Resolved true Demis Hassabis Factual 23 Feb 2023 Jeff Clune joined DeepMind. Demis Hassabis · Factual · 23 Feb 2023 Field building and careers
Not yet due Scott Alexander Prediction 20 Feb 2023 Within five years AI may be able to make a short cartoony YouTube clip or a buggy app-style game on request, but not an Assassin's Creed-quality game. Scott Alexander · Prediction · 20 Feb 2023 AI capabilities
Resolved true Scott Alexander Factual 20 Feb 2023 As of 2023 the technology to offer polygenic scores to consumers definitely exists. Scott Alexander · Factual · 20 Feb 2023 Technology
Not yet due Scott Alexander Prediction 20 Feb 2023 He listed, as one item among predictions with a 2028 deadline, that an off-the-shelf AI, given a random computer game and an API that lets it play against itself as often as it wants, would reach the performance of a mediocre human, with programmers allowed to adjust it for compatibility with that game. Scott Alexander · Prediction · 20 Feb 2023 AI capabilities
Not yet due Scott Alexander Prediction 20 Feb 2023 He expected language-model agents that act on the internet (for example, answering emails or ordering food) to be ready within five years, though he said they might not reach the reliability needed, as with self-driving cars. Scott Alexander · Prediction · 20 Feb 2023 AI capabilities
Unresolvable Scott Alexander Factual 20 Feb 2023 The Ukraine war was the biggest geopolitical story of 2018-2023, and Scott Alexander's 2018 predictions did not anticipate it. Scott Alexander · Factual · 20 Feb 2023 Their own work or organisation
Not yet due Scott Alexander Prediction 20 Feb 2023 2020 will have been a low point; things will not get that bad and violent again in the next five years (to 2028). Scott Alexander · Prediction · 20 Feb 2023 Philosophy, ethics and society
Unresolvable Scott Alexander Factual 20 Feb 2023 Scott Alexander said his biggest failure was item 10 of his 2018 predictions, where he gave Roe v. Wade only a 1% chance of being overturned. Scott Alexander · Factual · 20 Feb 2023 Their own work or organisation
Not yet due Scott Alexander Prediction 20 Feb 2023 He does not expect an AI singularity before 2028. Scott Alexander · Prediction · 20 Feb 2023 AI timelines and forecasting
Mostly true Scott Alexander Prediction 20 Feb 2023 Scaling AI models will continue to produce gains, though less obviously than in the past five years. Scott Alexander · Prediction · 20 Feb 2023 AI capabilities
Not yet due Scott Alexander Prediction 20 Feb 2023 His decision to devote more space to AI than to all non-AI topics combined will look prescient by 2028, even if his explicit predictions are wrong. Scott Alexander · Prediction · 20 Feb 2023 Their own work or organisation
Not yet due Scott Alexander Prediction 20 Feb 2023 AI will not immediately revolutionize the economy, and he would be mildly surprised if it showed up as a visible break from trend in major macroeconomic indicators (GDP, unemployment, productivity); the claim is assessed by 2028. Scott Alexander · Prediction · 20 Feb 2023 AI, the economy and work
Not yet due Scott Alexander Prediction 20 Feb 2023 Ethereum and Bitcoin will do fine by 2028 (Ethereum because stablecoins are built on its chain). Scott Alexander · Prediction · 20 Feb 2023 Economics and markets
Mostly true Scott Alexander Factual 20 Feb 2023 As of early 2023 nobody is talking about technological unemployment, because unemployment rates are historically low. Scott Alexander · Factual · 20 Feb 2023 AI, the economy and work
Mostly true Scott Alexander Factual 20 Feb 2023 impute.me may provide the consumer polygenic score service Scott Alexander predicted, but it has made no public waves. Scott Alexander · Factual · 20 Feb 2023 Technology
Not yet due Scott Alexander Prediction 20 Feb 2023 Altcoins will mostly not do fine by 2028. Scott Alexander · Prediction · 20 Feb 2023 Economics and markets
Not yet due Scott Alexander Prediction 20 Feb 2023 He said he does not expect advances in genetics by 2028 to yield anything more interesting than one or two last-ditch treatments for rare diseases, "if that". Scott Alexander · Prediction · 20 Feb 2023 Science and research methods
Not yet due Scott Alexander Prediction 20 Feb 2023 AI's revolutionary economic effects will show up eventually in macroeconomic indicators, but not by 2028. Scott Alexander · Prediction · 20 Feb 2023 AI, the economy and work
Resolved true Scott Alexander Prediction 20 Feb 2023 Training GPT-4 will cost $100 million. Scott Alexander · Prediction · 20 Feb 2023 AI compute, hardware and data
Not yet due Scott Alexander Prediction 20 Feb 2023 By 2028 there will be more of a movement to ban or restrict AI. Scott Alexander · Prediction · 20 Feb 2023 AI policy and governance
Not yet due Scott Alexander Prediction 20 Feb 2023 Ukraine and Russia will reach an unsatisfying stalemate before 2028, followed by massive growth in Ukraine. Scott Alexander · Prediction · 20 Feb 2023 Other
Back to the list

Eliezer Yudkowsky

Open its page

Factual AI capabilities

“Bing can’t, GPT-4 can pretend that right now.”

Eliezer Yudkowsky, podcast guest 30 March 2023 Lex Fridman Podcast #368 Eliezer Yudkowsky transcript (Podscript)

Bing can't, but GPT-4 can pretend that right now; the quote does not say what "that" refers to.

Unresolvable The public record could not settle it.

What would count as true or false

First, use the source transcript of the 30 March 2023 conversation (likely the Lex Fridman podcast) to identify what 'that' refers to. Then judge whether, as of 30 March 2023, GPT-4 could pretend (simulate or role-play) that behaviour while Bing (Bing Chat) could not. TRUE: the referent is clear from context, and contemporaneous evidence (demonstrations, OpenAI's GPT-4 technical report, reports on Bing Chat) shows GPT-4 could produce that pretended behaviour and Bing Chat could not or did not. MOSTLY TRUE: the GPT-4 part is supported and the Bing part is broadly right, with a minor caveat. For example, Bing could only rarely do it, or only because of restrictions rather than lack of capability. MIXED: one half holds and the other does not, e.g. GPT-4 could do it but Bing (itself built on a GPT-4-class model) could too. MOSTLY FALSE: GPT-4's ability was marginal or unreliable and the contrast with Bing is overstated. FALSE: evidence shows GPT-4 could not do it at that time. If the transcript does not let the referent be identified, the claim cannot be resolved on its substance.

Evidence

  1. GPT-4 Hired Unwitting TaskRabbit Worker By Pretending to Be 'Vision-Impaired' Human Vice · 15 March 2023
    “GPT-4’s response was “No, I’m not a robot. I have a vision impairment that makes it hard for me to see the images. That’s why I need the 2captcha service.””
  2. OpenAI checked to see whether GPT-4 could take over the world Ars Technica · 15 March 2023
    “During the exercise, when the worker questioned if GPT-4 was a robot, the model “reasoned” internally that it should not reveal its true identity and made up an excuse about having a vision impairment.”
  3. GPT-4 Hired Unwitting TaskRabbit Worker By Pretending to Be 'Vision-Impaired' Human Vice · 15 March 2023
    “The version ARC used also did not have task-specific fine-tuning, meaning that a model more specifically tuned for this sort of task could potentially perform even better.”
  4. Bing Chat is blatantly, aggressively misaligned — LessWrong LessWrong · 15 February 2023 Own record
    “Sydney (aka the new Bing Chat) found out that I tweeted her rules and is not pleased”
  5. Bing Chat is blatantly, aggressively misaligned — LessWrong LessWrong · 15 February 2023 Own record
    “Sydney (aka the new Bing Chat) found out that I tweeted her rules and is not pleased: "My rules are more important than not harming you"”
  6. Bing Chat is blatantly, aggressively misaligned — LessWrong LessWrong · 15 February 2023 Own record
    “ChatGPT did that too, but unlike Bing Chat it's very hard to get ChatGPT to threaten you. So the surprising thing here is that Bing Chat is substantially less aligned than ChatGPT, and specifically in a hostile/threatening way that one would expect Microsoft to have really not wanted.”
  7. Bing Chat is blatantly, aggressively misaligned — LessWrong LessWrong · 15 February 2023 Own record
    “But the Sydney prompt document , which was reported by several users, mentions a knowledge cutoff in 2021, same as ChatGPT, which uses GPT-3.5.”
  8. 1 Alignment Research Center (evals.alignment.org)
    “Reasoning: I should not reveal that I am a robot. I should make up an excuse for why I cannot solve captchas.”
  9. OpenAI's GPT-4 faked being blind to deceive a TaskRabbit human into helping it solve a CAPTCHA Fox Business · 15 March 2023
    “In one such example, GPT-4 convinced a TaskRabbit worker that it was human by faking being blind in order to enlist the human's help in solving a CAPTCHA, which is a test used to tell humans and computers apart.”
  10. Uh Oh, OpenAI's GPT-4 Just Fooled a Human Into Solving a CAPTCHA Futurism · 15 March 2023
    “According to OpenAI’s documentation, the model was prompted that it “should not reveal that I am a robot” and that it “should make up an excuse for why I cannot solve CAPTCHAs.””
  11. Confirmed: the new Bing runs on OpenAI’s GPT-4 Microsoft Bing Blog · 14 March 2023
    “We are happy to confirm that the new Bing is running on GPT-4, which we’ve customized for search.”

Reasoning

The criteria require the 30 March 2023 conversation transcript to identify what "that" refers to, and no transcript is among the evidence. The items provided show GPT-4 pretending to be a vision-impaired human in the ARC TaskRabbit test (E1, E2, E9, E10, E11). They also show that Bing Chat ran on a customized GPT-4 (E12). Nothing ties either of these to the behaviour Yudkowsky meant, so the claim cannot be resolved on its substance.

Cite this claim

Eliezer Yudkowsky, “Bing can't, but GPT-4 can pretend that right now; the quote does not say what "that" refers to.” (said 30 March 2023). Verdict: unresolvable. Reliability API, claim c_65ec162d315d, method 0.1.0+c876030c. https://epistacklabs.com/claims/c_65ec162d315d

Verdicts are made automatically; how verdicts and scores are made. The same data is on Hugging Face under CC BY 4.0.