ClaimsYann LeCun

“If it were the case, we would have AI systems that could teach themselves to drive a car in 20 hours of practice, like any 17 year-old.”

Yann LeCun, on X 14 March 2024 Post by @ylecun on X, 2024-03-14 (X)

Factual AI capabilities

There are currently no AI systems that can teach themselves to drive a car in about 20 hours of practice, as a 17-year-old can.

Mostly true True in substance, wrong in a detail.

Factual: Statements about the present or the past, checked against the record: did they hold up?

The second judge would not resolve this claim; the verdict shown is the first judge's.

What would count as true or false

Judged as of 14 March 2024. 'Drive a car' means general driving competence comparable to a newly licensed teenager on real roads, not a narrow subtask. TRUE: no documented AI system had, by that date, learned general real-world driving from roughly 20 hours (order of tens of hours) of its own practice; existing autonomous-driving systems relied on far more data or engineering. MOSTLY TRUE: the claim holds for general driving, but some system had learned a narrow driving task (e.g., lane-following) in that much practice. MOSTLY FALSE: a system had learned near-general road driving from somewhat more practice, or with substantial pre-training that is debatable as 'teaching itself'. FALSE: a documented system had by then taught itself general real-world driving in about 20 hours of practice.

Set before the research began.

Evidence (4)

  1. E1 After just 20 hours of training, Wayve's fast-learning AI car is already driving itself on unfamiliar roads New Atlas · 4 April 2019

    “British startup Wayve taught a car to teach itself to drive, and using only some cameras, a sat-nav and 20 hours' worth of experience, it's already driving itself short distances on unfamiliar UK roads.”
  2. E2 Learning to drive like a human Wayve · 3 April 2019

    “What is more remarkable, is that this model learned this driving behaviour with only 20 hours of training data.”
  3. E3 Learning to drive like a human Wayve · 3 April 2019

    “At all times, trained safety drivers ensure the AI system obeys the road code safely.”
  4. E4 Learned urban driving Wayve · 2 December 2019

    “As a result of this additional learning signal, we were able to learn a useful driving policy for simple urban driving with 30 hours of labelled control data.”

Reasoning

E1 and E2 report that Wayve's 2019 system learned to drive itself short distances on unfamiliar UK roads from 20 hours of training data. That is the scale the claim names, but it is limited driving, not general competence like a newly licensed teenager. E3 notes that trained safety drivers ensured the system obeyed the road code at all times, and E4 describes only "simple urban driving" learned from 30 hours of labelled control data. Under the criteria, a system that learned a narrow or limited driving task in tens of hours, with general real-world driving still not achieved, makes the claim mostly true.

Cite this claim

Yann LeCun, “There are currently no AI systems that can teach themselves to drive a car in about 20 hours of practice, as a 17-year-old can.” (said 14 March 2024). Verdict: mostly true. Reliability API, claim c_f57310e1d8cb, method 0.1.0+c876030c, dataset release v1.1.2. https://epistacklabs.com/claims/c_f57310e1d8cb

This claim is in dataset release v1.1.2, with the same verdict.

Reply to this claim

For Yann LeCun, or someone speaking for them: a correction or counter-evidence. Replies are read before anything is shown; an approved reply shows your name, role, date, statement and link. Your email address is never shown and is deleted 30 days after the decision (privacy).

More factual claims by Yann LeCun

  1. Mostly true Factual 18 March 2024
    “For starters, transformers are equivariant to permutations.”

    Transformers are equivariant to permutations, as ConvNets are equivariant to translations.

  2. Mostly true Factual 8 March 2024
    “The architecture is good, the architecture of the encoder is good, but the fact that you train the system to reconstruct images does not lead it to produce to long, good generic features of images.”

    Training an encoder to reconstruct images does not lead it to produce good generic image features, though the encoder architecture is good.

  3. Mostly true Factual 8 March 2024
    “LLMs can do none of those or they can only do them in a very primitive way and they don’t really understand the physical world.”

    Current LLMs can do none of understanding the physical world, persistent memory, reasoning and planning, or only in a very primitive way.

  4. Mostly true Factual 8 March 2024
    “No, there’s one thing that autoregressive LLMs or that LLMs in general, not just the autoregressive one, but including the bird style bidirectional ones, are exploiting and its self supervised…”

    LLMs, including BERT-style bidirectional ones, exploit self-supervised learning, and LeCun has been a very strong advocate of self-supervised learning for many years.

  5. Mostly true Factual June 2023
    “However, masked AEs are less successful with images, so in continuous domains, in general, [56].”

    Masked autoencoders are less successful with images and other continuous domains than with discrete domains such as language.

All of them

Verdicts are made automatically; how verdicts are made.