TL;DR
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
Scott Alexander has published an open-letter response to Steven Pinker’s arguments that some AI-risk warnings rely on an oversimplified view of intelligence. Alexander says successive language-model generations have improved and argues there is no compelling basis for assuming progress will stop before reaching human-level capability. The supplied text covers only part of Alexander’s response, and it provides no evidence that Pinker has accepted a debate.
Scott Alexander, writing at Astral Codex Ten, has published an open letter responding to Steven Pinker’s arguments about AI risk, disputing Pinker’s view that warnings about advanced AI depend on treating intelligence as a single, steadily increasing quantity. Alexander argues that the progress between successive language-model generations makes it reasonable to expect further gains, while acknowledging that capability growth could eventually level off.
The letter follows an earlier exchange in which Alexander challenged Pinker to a public debate after becoming frustrated with Pinker’s social-media posts about AI. Pinker responded with an open letter of his own, according to Alexander, saying that in-person debates can feel confrontational and are not necessarily a good way to seek truth. Alexander apologizes for any unintended harassment caused by the debate proposal and says he will answer the arguments Pinker set out.
Alexander’s first substantive response addresses Pinker’s objection that AI-risk scenarios often treat intelligence as a general “quantity of power,” extending a scale from animals through humans to superintelligent AI. Pinker’s position, as quoted and summarized in Alexander’s letter, is that intelligent systems are mechanisms suited to some problems but not others, and are limited by what they can learn about the world through observation and experimentation.
Alexander counters that people can reasonably describe later GPT systems as more intelligent than earlier ones in ordinary usage. He points to examples of models improving at tasks that earlier systems struggled with, including basic arithmetic and harder problems. The letter also refers to an internal OpenAI model said to have solved the Navier-Stokes equations; that is Alexander’s account, and the supplied material gives no model name, technical details, or independent confirmation.
The Dispute Over AI Capability Growth
The exchange matters because it puts a central disagreement in AI-risk discussions into direct view: whether observed gains in specific tasks support expectations of broader, continuing increases in capability. Alexander’s case is that improvements across successive model generations weaken the assumption that current limitations will persist. He says no compelling argument establishes where progress will stop, particularly not just below human-level performance.
Pinker’s counterargument, as presented in the letter, challenges the idea that intelligence can be treated as a simple ladder. If an AI system’s abilities remain uneven and depend on information it can acquire, then improvements on selected tests do not by themselves establish general human-like competence or future superintelligence. The distinction affects how readers interpret model demonstrations: a system can improve substantially on some problems without showing that every relevant ability is advancing at the same rate.
The piece is an argumentative exchange, not a new technical study or a report of a new AI release. Its public value lies in clarifying the competing reasoning and claims made by two prominent writers. Readers should distinguish Alexander’s examples and predictions from independently established findings about how far AI systems can progress.
As an affiliate, we earn on qualifying purchases.
How the Open-Letter Exchange Began
Alexander says he initially proposed a public debate after taking issue with Pinker’s social-media comments on AI. Pinker declined the format, according to Alexander’s account, and instead wrote an open letter explaining his objections and setting out arguments against some AI-risk concerns. Alexander says he agrees that face-to-face debates can be confrontational, but describes his original proposal as a response to what he regarded as dishonorable public writing.
The response revisits earlier disputes over language-model capability. Alexander recalls a 2022 debate involving Scott Aaronson and Gary Marcus, in which Marcus cited basic arithmetic errors by GPT-2 and GPT-3. Alexander says he expected scaling to address such failures; he says GPT-4 handled the example and later systems performed harder tasks. He also cites a 2023 prediction by Pinker that improvement would not continue exponentially. These historical claims and quotations are presented through Alexander’s account of the earlier exchanges.
“GPT-6 is more intelligent than GPT-3.”
— Scott Alexander, in his open letter
As an affiliate, we earn on qualifying purchases.
Evidence Behind the Capability Claims
The supplied source text ends partway through Alexander’s discussion of research on general cognitive ability in animals, so it does not show the full argument or his complete response to Pinker’s other points. It is also unclear from the material whether Pinker has replied to this letter or whether either writer intends to continue the exchange.
Several technical examples need further documentation to assess independently. The letter does not identify the internal OpenAI model said to have solved Navier-Stokes, provide a paper or evaluation, or describe what “solved” means in that instance. Nor does the excerpt establish that performance gains on the cited tasks demonstrate general intelligence, predict future progress, or settle the possibility of AI risk.
As an affiliate, we earn on qualifying purchases.
Whether Pinker Responds Again
The immediate next development would be a response from Pinker, if he chooses to answer Alexander’s points. No further exchange, scheduled debate, or publication date is announced in the supplied material. Readers seeking to evaluate the technical claims will also need fuller details and independent evidence for the cited model results, alongside the complete letters from both writers.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Scott Alexander responding to?
He is answering an open letter from Steven Pinker that, as Alexander describes it, challenges aspects of AI-risk arguments and questions whether intelligence should be treated as a general scale that can be extrapolated to superintelligence.
Did Pinker agree to a public debate?
The supplied material does not say that Pinker accepted Alexander’s debate challenge. Alexander says Pinker responded with an open letter and objected to in-person debates as a confrontational format.
What is Alexander’s main argument in the excerpt?
He argues that successive GPT models have improved on tasks earlier versions handled poorly, and that there is no compelling basis for assuming progress will stop at a particular level, including just below human capability.
Does the letter establish that AI will reach human-level intelligence?
No. Alexander argues that further progress is a reasonable default expectation, but the excerpt does not prove that AI will reach human-level capability or establish when progress might level off.
What is unclear about the cited OpenAI result?
The excerpt says an internal model beyond GPT-6 solved Navier-Stokes, but gives no model name, test details, publication, or independent confirmation. The claim is attributed to Alexander’s letter.
Source: rss
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
