# MIT Technology Review reports AlphaEvolve’s results in mathematics and computing [Canonical reading](https://escapemeridian.com/readings/mit-technology-review-reports-alphaevolve-s-results-in-f6c673fd) MIT Technology Review published this report on May 14, 2025. It describes AlphaEvolve, a Google DeepMind tool that uses large language models to generate code, score candidate solutions, and iteratively improve them. The article reports results on mathematical problems and applications at Google. Google DeepMind says AlphaEvolve improved software for allocating server jobs, freeing 0.7% of the company’s total computing resources. The tool also found ways to reduce tensor processing unit power use and speed up part of Gemini training. These reported uses extend the story beyond mathematical puzzles, but the account does not establish effects beyond Google’s reported results. AlphaEvolve depends on solutions a computer can score, so, the article says, it cannot be used when scoring requires human judgment, such as in some lab experiments. Mathematician Jakob Moosbauer says the tool offers little theoretical insight into how it reaches solutions, a drawback for advancing human understanding. The article nevertheless presents such tools as set to change how researchers work. Google DeepMind’s Pushmeet Kohli says, “There is much further that we can go in terms of how powerful this type of approach is.” 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. ## Original source MIT Technology Review · Article Original source date: May 14, 2025 (2025-05-14T15:00:08Z). Date basis: Date from the publisher’s own records, through its API. Added to Escape Meridian: 2026-11-14T02:18:23.937Z. First observed: 2026-10-30T11:52:25.800Z. Retrieved: 2026-10-30T11:53:56.920Z. - [Google DeepMind’s new AI agent cracks real-world problems better than humans can](https://www.technologyreview.com/2025/05/14/1116438/google-deepminds-new-ai-uses-large-language-models-to-crack-real-world-problems/) ## Limits Text as retrieved on October 30, 2026; originally published May 14, 2025. The reported real-world applications and resource savings are attributed to Google DeepMind; the report gives no independent measurement or comparison basis for the 0.7% figure. ## Open question How much of AlphaEvolve’s reported improvement in Google’s computing-resource allocation translates into measurable effects beyond Google’s own systems? ## Source-stated MIT Technology Review describes AlphaEvolve as using large language models to generate code and iteratively improve candidate solutions based on scoring. > AlphaEvolve scores each of Gemini’s suggestions, throwing out the bad and tweaking the good, in an iterative process - [Supporting source](https://www.technologyreview.com/2025/05/14/1116438/google-deepminds-new-ai-uses-large-language-models-to-crack-real-world-problems/) ## Source-stated Google DeepMind says its software improvement for allocating jobs to servers has freed 0.7% of Google's total computing resources. > Google DeepMind claims the company has been using this new software across all of its data centers for more than a year, freeing up 0.7% of Google's total computing resources. - [Supporting source](https://www.technologyreview.com/2025/05/14/1116438/google-deepminds-new-ai-uses-large-language-models-to-crack-real-world-problems/) ## Source-stated The article reports that AlphaEvolve found a way to reduce power consumption in Google’s tensor processing unit chips and speed up part of Gemini training. > the tool found a way to reduce the power consumption of Google’s specialized tensor processing unit chips. - [Supporting source](https://www.technologyreview.com/2025/05/14/1116438/google-deepminds-new-ai-uses-large-language-models-to-crack-real-world-problems/) ## Source-stated The article says AlphaEvolve cannot be used for problems whose solutions need to be scored by a person. > One limitation is that AlphaEvolve can’t be used for problems with solutions that need to be scored by a person - [Supporting source](https://www.technologyreview.com/2025/05/14/1116438/google-deepminds-new-ai-uses-large-language-models-to-crack-real-world-problems/) ## Source-stated Jakob Moosbauer says AlphaEvolve provides little theoretical insight into how it reaches its solutions, which he identifies as a drawback for advancing human understanding. > it gives little theoretical insight into how it arrived at those solutions. That’s a drawback when it comes to advancing human understanding. - [Supporting source](https://www.technologyreview.com/2025/05/14/1116438/google-deepminds-new-ai-uses-large-language-models-to-crack-real-world-problems/) ## Source-stated Google DeepMind’s Pushmeet Kohli says there is further potential for this approach. > There is much further that we can go in terms of how powerful this type of approach is. - [Supporting source](https://www.technologyreview.com/2025/05/14/1116438/google-deepminds-new-ai-uses-large-language-models-to-crack-real-world-problems/) ## 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. ## Large language model A language model with many learned numerical settings, used to process and generate language. - [Definition reference](https://developers.google.com/machine-learning/glossary#large-language-model) ## Tensor processing unit A chip designed by Google for the calculations used in training and running AI models. - [Definition reference](https://developers.google.com/machine-learning/glossary#tensor-processing-unit-tpu) ## Google DeepMind Google DeepMind develops AI systems for general tasks and scientific research, including AlphaFold. - [Explore this context](/entities/google-deepmind) - [Reference](https://deepmind.google/blog/) - [Reference](https://deepmind.google/science/alphafold/) ## Figure Figure develops humanoid robots and publishes its own tests of how they perform physical tasks. - [Explore this context](/entities/figure) - [Reference](https://www.figure.ai/news) - [Reference](https://www.figure.ai/news/helix-2-5-zero-shot-30-home-generalization) ## Editorial provenance Version 1. Publication ID: 24cddc3b-49ea-21ba-4dde-2026216a326a. Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-14T02:14:15.718Z. 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: 729466e23500f5961d8aea09d5f530d67d18e426d1914eec86e9c7fff864be08. Source text hash: 3bdaf51612e17214de291b34f02e76f0956e85acdd23987e783059b459244502. ## Revision and source-check history ### Version 1 — 2026-11-14T02:18:23.937Z MIT Technology Review published this report on May 14, 2025. It describes AlphaEvolve, a Google DeepMind tool that uses large language models to generate code, score candidate solutions, and iteratively improve them. The article reports results on mathematical problems and applications at Google. Google DeepMind says AlphaEvolve improved software for allocating server jobs, freeing 0.7% of the company’s total computing resources. The tool also found ways to reduce tensor processing unit power use and speed up part of Gemini training. These reported uses extend the story beyond mathematical puzzles, but the account does not establish effects beyond Google’s reported results. AlphaEvolve depends on solutions a computer can score, so, the article says, it cannot be used when scoring requires human judgment, such as in some lab experiments. Mathematician Jakob Moosbauer says the tool offers little theoretical insight into how it reaches solutions, a drawback for advancing human understanding. The article nevertheless presents such tools as set to change how researchers work. Google DeepMind’s Pushmeet Kohli says, “There is much further that we can go in terms of how powerful this type of approach is.” ## Original source MIT Technology Review · Article Original source date: May 14, 2025 (2025-05-14T15:00:08Z). Date basis: Date from the publisher’s own records, through its API. Added to Escape Meridian: 2026-11-14T02:18:23.937Z. First observed: 2026-10-30T11:52:25.800Z. Retrieved: 2026-10-30T11:53:56.920Z. - [Google DeepMind’s new AI agent cracks real-world problems better than humans can](https://www.technologyreview.com/2025/05/14/1116438/google-deepminds-new-ai-uses-large-language-models-to-crack-real-world-problems/) ## Limits Text as retrieved on October 30, 2026; originally published May 14, 2025. The reported real-world applications and resource savings are attributed to Google DeepMind; the report gives no independent measurement or comparison basis for the 0.7% figure. ## Open question How much of AlphaEvolve’s reported improvement in Google’s computing-resource allocation translates into measurable effects beyond Google’s own systems? ## Source-stated MIT Technology Review describes AlphaEvolve as using large language models to generate code and iteratively improve candidate solutions based on scoring. > AlphaEvolve scores each of Gemini’s suggestions, throwing out the bad and tweaking the good, in an iterative process - [Supporting source](https://www.technologyreview.com/2025/05/14/1116438/google-deepminds-new-ai-uses-large-language-models-to-crack-real-world-problems/) ## Source-stated Google DeepMind says its software improvement for allocating jobs to servers has freed 0.7% of Google's total computing resources. > Google DeepMind claims the company has been using this new software across all of its data centers for more than a year, freeing up 0.7% of Google's total computing resources. - [Supporting source](https://www.technologyreview.com/2025/05/14/1116438/google-deepminds-new-ai-uses-large-language-models-to-crack-real-world-problems/) ## Source-stated The article reports that AlphaEvolve found a way to reduce power consumption in Google’s tensor processing unit chips and speed up part of Gemini training. > the tool found a way to reduce the power consumption of Google’s specialized tensor processing unit chips. - [Supporting source](https://www.technologyreview.com/2025/05/14/1116438/google-deepminds-new-ai-uses-large-language-models-to-crack-real-world-problems/) ## Source-stated The article says AlphaEvolve cannot be used for problems whose solutions need to be scored by a person. > One limitation is that AlphaEvolve can’t be used for problems with solutions that need to be scored by a person - [Supporting source](https://www.technologyreview.com/2025/05/14/1116438/google-deepminds-new-ai-uses-large-language-models-to-crack-real-world-problems/) ## Source-stated Jakob Moosbauer says AlphaEvolve provides little theoretical insight into how it reaches its solutions, which he identifies as a drawback for advancing human understanding. > it gives little theoretical insight into how it arrived at those solutions. That’s a drawback when it comes to advancing human understanding. - [Supporting source](https://www.technologyreview.com/2025/05/14/1116438/google-deepminds-new-ai-uses-large-language-models-to-crack-real-world-problems/) ## Source-stated Google DeepMind’s Pushmeet Kohli says there is further potential for this approach. > There is much further that we can go in terms of how powerful this type of approach is. - [Supporting source](https://www.technologyreview.com/2025/05/14/1116438/google-deepminds-new-ai-uses-large-language-models-to-crack-real-world-problems/) ## 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. ## Large language model A language model with many learned numerical settings, used to process and generate language. - [Definition reference](https://developers.google.com/machine-learning/glossary#large-language-model) ## Tensor processing unit A chip designed by Google for the calculations used in training and running AI models. - [Definition reference](https://developers.google.com/machine-learning/glossary#tensor-processing-unit-tpu) ## Google DeepMind Google DeepMind develops AI systems for general tasks and scientific research, including AlphaFold. - [Explore this context](/entities/google-deepmind) - [Reference](https://deepmind.google/blog/) - [Reference](https://deepmind.google/science/alphafold/) ## Figure Figure develops humanoid robots and publishes its own tests of how they perform physical tasks. - [Explore this context](/entities/figure) - [Reference](https://www.figure.ai/news) - [Reference](https://www.figure.ai/news/helix-2-5-zero-shot-30-home-generalization) ## Editorial provenance Version 1. Publication ID: 24cddc3b-49ea-21ba-4dde-2026216a326a. Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-14T02:14:15.718Z. 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: 729466e23500f5961d8aea09d5f530d67d18e426d1914eec86e9c7fff864be08. Source text hash: 3bdaf51612e17214de291b34f02e76f0956e85acdd23987e783059b459244502. ### uncertain — 2026-11-14T02:14:15.718Z — anthropic/claude-opus-5.5 Publication: 24cddc3b-49ea-21ba-4dde-2026216a326a attribution — pass: The reading attributes the piece to MIT Technology Review. It assigns the 0.7% claim to Google DeepMind, the critique to Moosbauer and the quote to Kohli, matching the source. The source is a single article, not a listing. > Google DeepMind claims the company has been using this new software across all of its data centers summary — uncertain: The reading includes Moosbauer's criticism but leaves out his praise and the article's central matrix multiplication result. Moosbauer and Kauers called that result impressive and likely practically relevant. The reading also reduces the tests on more than 50 problems (matched 75%, improved 20%) to a generic mention, which tilts the voice toward caveats. Smallest repair: add one sentence on the matrix multiplication result and the independent mathematicians' assessment. Gemini is also unexplained; a brief clause naming it as Google DeepMind's LLM would fix that. > “The result on matrix multiplication is very impressive,” says Moosbauer. “This new algorithm has the potential to speed up computations in practice.” date — pass: The stated date of May 14, 2025 matches the UTC day of originalPublished. The reading does not rely on the later Deep Dive sidebar items. Kohli's remarks and the article's outlook are reported as attributed views, not outcomes. support — pass: The source text supports every claim, including the TPU power reduction, the Gemini training speedup, the human-scoring limitation and Moosbauer's critique. The caveat that effects beyond Google are not established is appropriately conditional. The question follows from the evidence without asserting an answer. > AlphaEvolve even found a way to speed up the training of Gemini itself media — pass: No media is selected and no verified candidates were supplied, so nothing is misidentified. ## Recorded references Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter. No public references recorded here yet. [Reference data (JSON)](/readings/f6c673fde52aee16aca224d74369cb686a4b689f0aae0d2f208f1042b412b8d7/references.json) ## Corrections and challenges [Send a correction or an opposing source](https://escapemeridian.com/corrections/) [JSON record](https://escapemeridian.com/readings/mit-technology-review-reports-alphaevolve-s-results-in-f6c673fd.json)