What’s wrong with Google’s new robot project
We recorded 20 claims from this post. 8 were checked against evidence. 12 can’t be checked. 4 more are held. Read the post at its source.
Claims from this post (20)
In the order they appear in the post. Each quotation is the speaker’s words in this post. A label such as “Supported” or “Can’t be checked” is the claim’s current assessment; its page has the evidence and the reasons.
Assessment: Can’t be checkedSaid
“I am not surprised that this can (kinda sorta) be done, but I am not sure it should be done.”
What we tested: Marcus doubts that language-model-driven robots like PaLM-SayCan should be built, even though they can sort of be made to work.
Assessment: SupportedSaid
“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.”
What we tested: The language technology that PaLM-SayCan relies on is well known to be problematic.
What happened: PaLM's own announcement acknowledges an existing research literature on risks and describes analyses of bias and model outputs. This directly supports recognition of problems in the relevant technology. (Google Research, )
Assessment: Can’t be checkedSaid
“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.”
What we tested: Language model technology is likely to be even more problematic when used in robots.
Assessment: Can’t be checkedSaid
“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.”
What we tested: Large language models are powerful but reckless, like bulls in a china shop.
Assessment: Can’t be checkedSaid
“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.”
What we tested: If large language models are put inside robots and misunderstand a request or miss its implications, they could cause major damage.
Assessment: Can’t be checkedSaid
“The PaLM-SayCam experiment makes it pretty clear that the list of 21 problems is due for an update.”
What we tested: DeepMind's list of 21 ethical and social problems with LLMs should be updated to cover the risks of putting LLMs in robots.
Assessment: Can’t be checkedSaid
“Misunderstandings will arise; without really firm checks on these systems, some of those misunderstandings could lead to mayhem.”
What we tested: LLM-driven robots will misunderstand people, and without firm checks some of those misunderstandings could lead to mayhem.
Assessment: UnresolvedSaid
“Building a robot on top of language system that has such little comprehension of the world can’t be a recipe for success.”
What we tested: Building a robot on top of a language system with so little comprehension of the world cannot be a recipe for success.
Assessment: Can’t be checkedSaid
“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.”
What we tested: 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.
Assessment: Can’t be checkedSaid
“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.”
What we tested: The AI field currently has no compelling idea of how to solve the alignment problems of large language models.
Assessment: Can’t be checkedSaid
“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.”
What we tested: Google's PaLM-SayCan stitches superficial, incorrigible language models onto powerful and potentially dangerous humanoid robots.
Assessment: SupportedSaid
“Driverless car demos have been around for decades, but making them work reliably has proven far harder than most people realized.”
What we tested: Making driverless cars work reliably has proven far harder than most people realized.
What happened: The report records a direct industry admission and describes executives narrowing expectations because of technical complexity. A readable primary transcript was not obtained, so contemporary reporting is used. (WIRED, )
Assessment: UnresolvedSaid
“Google makes a big deal about how the system is to do some degree interpretable, but there is no obvious way to integrate large language models with the kind of formal formal verification methods that are now routine in the design of microprocessors, USB drivers and large aircraft.”
What we tested: There is no obvious way to integrate large language models with the kind of formal verification methods that are now routine in designing microprocessors, USB drivers and large aircraft.
Assessment: Can’t be checkedSaid
“Without a much greater degree of interpretability and verifiability, its hard to see how we can get to that level of safety.”
What we tested: 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.
Assessment: UnresolvedSaid
“Big data won’t be enough to solve the problems of robotics”
What we tested: More data alone will not be enough to solve the problems of robotics.
Assessment: UnresolvedSaid
“Facebook M was canceled, and 7 years later nobody yet can buy a super-smart (or even decently smart) robot at any price.”
What we tested: As of August 2022, nobody can buy even a decently smart robot at any price.
Assessment: Can’t be checkedSaid
“And the more actions a robot might consider, the worse performance is likely to be.”
What we tested: The more actions a robot like PaLM-SayCan has to consider, the worse its performance is likely to be.
Assessment: SupportedSaid
“Performance hovered at around 75% in the best case, when the robots had a limited number of options to consider.”
What we tested: PaLM-SayCan's performance was around 75% at best, when the robot had a limited number of options.
What happened: The reported overall execution rates are 74% in the mock kitchen and 60% in the real kitchen. Thus the better overall setting is approximately 75%; planning is a separate measure. (arXiv; Michael Ahn and colleagues, )
Assessment: UnresolvedSaid
“Cool demo, yes. Real world product, anytime soon? Fuggedaboutit.”
What we tested: PaLM-SayCan will not become a real-world product anytime soon.
Assessment: Can’t be checkedSaid
“In a general-purpose, humanoid domestic robot, 75% is nowhere near good enough.”
What we tested: In a general-purpose humanoid domestic robot, a 75% success rate is nowhere near good enough.
Held from this post (4)
These records did not pass every check, so nothing about them is published. Held is not a finding.
- The record does not meet the current rules. Our reading, not assessed: No one, including Elon Musk, should hold their breath expecting systems like PaLM-SayCan's vision, robots we can talk to that help with everyday chores, to appear in reliable form in the next several years.
- The record does not meet the current rules. Our reading, not assessed: There is no reason to think that robots, including robots with natural language interfaces like Google's new system, are exempt from the problems of infinite edge cases that deep nets easily forget.
- The two reviews reached different assessments. Our reading, not assessed: On some actions PaLM-SayCan executed correctly 0% of the time, which is inadequate for a household robot.
- The two reviews reached different assessments. Our reading, not assessed: Edge cases remain the core problem for driverless cars, and the industry has only recently recognized this.