Escape Meridian

MIT Technology Review · Mar 4, 2024

MIT Technology Review reports that deep learning still puzzles its researchers

Markdown · JSON · Original source

MIT Technology Review published this article on March 4, 2024. It argues that ’s success has outpaced understanding of why it works, and that this gap matters for developing future AI and anticipating risks. The article describes grokking: models may appear unable to learn a task, then succeed after more training. It also examines disagreement about double descent, a reported pattern in which error rates fall, rise, and fall again as models grow. Alicia Curth and colleagues argue that the pattern may be an illusion caused by how model complexity was measured; with another measure, large models might fit classical statistics. The article does not resolve this dispute. Researchers use simpler models to develop theories, but it is unclear whether those theories apply to large models, where many surprising behaviors occur. Without a better account of what models can do, the article argues, anticipating their capabilities and risks is difficult. Researchers do not know what capabilities GPT-5 will have until they train and test it.

Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.

MIT Technology ReviewArticle · technologyreview.comOpen original source

Original source

MIT Technology Review · Article

Source details

Original source date: Mar 4, 2024 (2024-03-04T10:51:41Z).

Date basis: Date from the publisher’s own records, through its API.

Added to Escape Meridian: 2026-11-21T09:08:10.018Z. First observed: 2026-10-28T16:36:21.982Z. Retrieved: 2026-10-28T16:37:55.827Z.

Referenced by

Recorded citations in public readings. A reference alone does not establish agreement or influence.

No public references recorded here yet.

Reference data (JSON)

Limits

Text as retrieved on October 28, 2026; originally published March 4, 2024.

Open question

Can theories developed with simpler models explain the surprising behaviors of large models?

Source-stated

MIT Technology Review’s article argues that understanding why deep learning works could help researchers develop future AI and anticipate risks.

figuring out why deep learning works so well isn’t just an intriguing scientific puzzle. It could also be key to unlocking the next generation of the technology—as well as getting a handle on its formidable risks.

Source-stated

The article describes grokking as cases where a model appears not to learn a task, then succeeds, and notes that researchers do not agree on what explains it.

models could fail to learn a task and then all of a sudden just get it, as if a lightbulb had switched on. This wasn’t how deep learning was supposed to work. They called the behavior grokking.

Source-stated

The article reports competing explanations of double descent: some researchers treat it as a challenge to classical statistics, while Alicia Curth and colleagues argue that it may result from how model complexity was measured.

At the same conference, Alicia Curth, who studies statistics at the University of Cambridge, and her colleagues argued that double descent is in fact an illusion.

Source-stated

The article says theories built using simpler models may not explain large models, where many unusual behaviors occur.

This helps get new theories off the ground, but it is not always clear if those theories will hold for larger models too. After all, it is in the complexity of large models that many of the weird behaviors reside.

Source-stated

The article says researchers cannot know what capabilities GPT-5 will have until it is trained and tested.

We don’t know what capabilities GPT-5 will have until we train it and test it

Source and contributor context

MIT Technology Review · Publisher

The publisher is identified here. No individual author has been inferred.

Source media

Media retrieval status: unchecked.

Terms and references

Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.

Deep learning

Machine learning that uses many layers of simple calculations; the basis of most current AI.

Editorial provenance

Version 1. Publication ID: 5d030720-605f-c91b-7250-3190923a507d.

Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.

Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-21T09:03:34.198Z. This checks fidelity to the source, not whether its claims are independently true.

Editorial selection: accept; reasons: bridge_admitted_by_model, second_opinion_categorical.

Observation: 43755623764edef3224ba1093768ad360d0d64632370b4745ba8431de1987878. Source text hash: 9f85aed8089cf603f2f0b70ad0795593413815aecbf160978bf7f278a6e76cd5.

Revision and source-check history

Version 1 · 2026-11-21T09:08:10.018Z

MIT Technology Review reports that deep learning still puzzles its researchers

MIT Technology Review published this article on March 4, 2024. It argues that ’s success has outpaced understanding of why it works, and that this gap matters for developing future AI and anticipating risks. The article describes grokking: models may appear unable to learn a task, then succeed after more training. It also examines disagreement about double descent, a reported pattern in which error rates fall, rise, and fall again as models grow. Alicia Curth and colleagues argue that the pattern may be an illusion caused by how model complexity was measured; with another measure, large models might fit classical statistics. The article does not resolve this dispute. Researchers use simpler models to develop theories, but it is unclear whether those theories apply to large models, where many surprising behaviors occur. Without a better account of what models can do, the article argues, anticipating their capabilities and risks is difficult. Researchers do not know what capabilities GPT-5 will have until they train and test it.

Original source

MIT Technology Review · Article

Source details

Original source date: Mar 4, 2024 (2024-03-04T10:51:41Z).

Date basis: Date from the publisher’s own records, through its API.

Added to Escape Meridian: 2026-11-21T09:08:10.018Z. First observed: 2026-10-28T16:36:21.982Z. Retrieved: 2026-10-28T16:37:55.827Z.

Limits

Text as retrieved on October 28, 2026; originally published March 4, 2024.

Open question

Can theories developed with simpler models explain the surprising behaviors of large models?

Source-stated

MIT Technology Review’s article argues that understanding why deep learning works could help researchers develop future AI and anticipate risks.

figuring out why deep learning works so well isn’t just an intriguing scientific puzzle. It could also be key to unlocking the next generation of the technology—as well as getting a handle on its formidable risks.

Source-stated

The article describes grokking as cases where a model appears not to learn a task, then succeeds, and notes that researchers do not agree on what explains it.

models could fail to learn a task and then all of a sudden just get it, as if a lightbulb had switched on. This wasn’t how deep learning was supposed to work. They called the behavior grokking.

Source-stated

The article reports competing explanations of double descent: some researchers treat it as a challenge to classical statistics, while Alicia Curth and colleagues argue that it may result from how model complexity was measured.

At the same conference, Alicia Curth, who studies statistics at the University of Cambridge, and her colleagues argued that double descent is in fact an illusion.

Source-stated

The article says theories built using simpler models may not explain large models, where many unusual behaviors occur.

This helps get new theories off the ground, but it is not always clear if those theories will hold for larger models too. After all, it is in the complexity of large models that many of the weird behaviors reside.

Source-stated

The article says researchers cannot know what capabilities GPT-5 will have until it is trained and tested.

We don’t know what capabilities GPT-5 will have until we train it and test it

Source and contributor context

MIT Technology Review · Publisher

The publisher is identified here. No individual author has been inferred.

Source media

Media retrieval status: unchecked.

Terms and references

Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.

Deep learning

Machine learning that uses many layers of simple calculations; the basis of most current AI.

Editorial provenance

Version 1. Publication ID: 5d030720-605f-c91b-7250-3190923a507d.

Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.

Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-21T09:03:34.198Z. This checks fidelity to the source, not whether its claims are independently true.

Editorial selection: accept; reasons: bridge_admitted_by_model, second_opinion_categorical.

Observation: 43755623764edef3224ba1093768ad360d0d64632370b4745ba8431de1987878. Source text hash: 9f85aed8089cf603f2f0b70ad0795593413815aecbf160978bf7f278a6e76cd5.

uncertain · 2026-11-21T09:03:34.198Z · anthropic/claude-opus-5.5

Publication: 5d030720-605f-c91b-7250-3190923a507d

attribution — uncertain: Most of the reading correctly attributes claims to MIT Technology Review and names Curth for the illusion argument. However, the summary's last sentence, "Researchers do not know what capabilities GPT-5 will have until they train and test it," is stated flatly, and the claim credits it to 'the article.' In the source, this is a quote from Lauro Langosco of Cambridge, given as a safety concern. Smallest repair: attribute it to Langosco, for example 'Langosco says researchers won't know...', and keep it in the 2024 frame.

“We don’t know what capabilities GPT-5 will have until we train it and test it,” says Langosco.

summary — pass: The title, summary and question faithfully capture the article's thesis: deep learning's success has outpaced understanding, and that gap matters for future development and risk. Grokking and double descent are explained in plain language, the Curth dispute is kept unresolved, and deep learning is defined beside the reading. The question about whether proxy-model theories transfer to large models follows directly from the source. The reading is somewhat flatter than the source's voice, which includes Belkin's 'alchemy' and Barak's steering-wheel line, but nothing essential is erased.

for all its runaway success, nobody knows exactly how—or why—it works.

date — pass: March 4, 2024 matches the UTC day of originalPublished (2024-03-04T10:51:41Z), stated at day precision. The Deep Dive sidebar refers to 'Ten years after AlphaGo' and is likely a later addition to the page, but the reading neither uses nor relies on it. Forecasts are not presented as outcomes. The GPT-5 sentence is a 2024 statement and should stay framed as the source's view at that time (see attribution). The source is a publisher page, so no archive disclosure is needed.

support — pass: Each claim is grounded in the main article body, not later page additions, despite the page having been edited after publication. Supported points: the risk/next-generation thesis; the grokking description; the lack of consensus ('I don't think there's a consensus'); Curth's argument that double descent is an illusion caused by complexity measurement; and the limits of theories built on simpler proxy models. No unsupported benefits, costs or forecasts are added.

Her team argued that the double-descent phenomenon—where models appear to perform better, then worse, and then better again as they get bigger—arises because of the way the complexity of the models was measured.

media — pass: No media was selected, and no verified candidates were supplied. Leaving media empty is appropriate.

Challenge or correct this reading

Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.