Oracleの冷酷な午前6時の解雇メールが、新たな削減の波の中でスタッフを襲う

警察がFlockカメラ映像を使用して無実の女性を13日間投獄: 「人生が終わったと思った」 Oracleの冷酷な午前6時の解雇メールが、新たな削減の波の中でスタッフを襲う: 「今日はあなたの最後の勤務日です」 Samsung、Dua Lipaは1500万ドルの画像訴訟で勝訴の可能性なし: 「テレビを支持したという証拠は何もない」 OpenAI、Hugging Face以前に2度目の不正AIインシデントがあったことを明らかに: 「もっと」あるかもしれない ## 記事一覧 [1] タイトル: GPT-5.6 Luna vs. GPT-6 Astra: Is a $1.20 Model Good Enough for Code Review? URL: https://entelligence.ai/blogs/gpt-5.6-luna-vs-gpt-6-astra-is-a-1.20-model-good-enough-for-code-review 本文: GPT-5.6 Luna vs GPT-6 Astra: Is a $1.20 Model Good Enough for Code Review?Aditya JhaSep 14, 2026Aditya JhaSep 14, 2026GPT-5.6 Luna costs $0.20 per million input tokens and $1.20 per million output tokens. GPT-6 Astra costs $10 and $50. On the same pull requests, one Luna review cost $0.0041 and one Astra review cost $0.113, a 28x difference.Our last post compared Astra with GPT-5.6 Sol. This time we wanted to know what you give up if every pull request goes through the cheapest model.The short answerLuna found 69 verified bugs across 50 pull requests. Astra found 92. Luna cost $0.20 for the whole run and Astra cost $5.66. Luna was wrong more often, with 24 of its 93 findings failing verification against Astra's 4 of 96, and it found 9 of the 24 security bugs where Astra found 19.Our read: Luna is good enough for everyday correctness bugs at that price, and we wouldn't let it review authentication or permission code on its own.How we ran itWe reused the setup from the Astra vs Sol post so the numbers line up.The pull requests are the 50 public benchmark PRs in the AI-Code-Review-Evals organization, ten each from Cal.com, Sentry, Discourse, Keycloak and Grafana. Each one introduces defects against a clean base branch.Luna and Astra got the same prompt on the same diffs. The prompt asks for correctness, security, concurrency, resource and error-handling bugs, and excludes style, naming, docs and test suggestions. Each model returned structured findings.Verification works the same way as before. For every pull request, the findings from Astra, Sol, Luna and the public Entelligence reviewer comments go into one anonymized list. GPT-6 Astra and GPT-5.6 Sol each judge that list separately against the diff, grouping duplicates and deciding whether each issue is a real bug. An issue only counts as verified when both judges call it real. They agreed on 91% of findings, and 143 distinct bugs passed both.Adding Luna's findings changed the pool the judges saw, so everything was judged again. Astra's verified count moved from 91 in the last post to 92 here, and Sol's from 107 to 108. Astra is also one of the two judges, which could favor it slightly. The limits section covers that.The resultsGPT-5.6 LunaGPT-6 AstraVerified bugs6992Findings raised9396Precision74%96%Total cost, 50 PRs$0.20$5.66Cost per verified bug$0.0030$0.061Mean time per review23s36sMean output tokens per review2,104688Luna found 75% as many verified bugs as Astra for 3.6% of the money. Per verified bug, Astra cost 20x more.Luna wrote 3.1x as many output tokens per review as Astra and still came in far cheaper, because its output price is 42x lower. It was also faster, at 23 seconds per review against 36.Your team would feel the precision gap first. About one Luna comment in four was wrong, while Astra was wrong 4 times in 96. Developers who already skim AI review comments will skim harder when a quarter of them are noise.Where Luna falls behindReaders of the last post asked us to split results by codebase and by bug type, because an overall score can hide a model that does well on one repository and badly on another. On this data, the split shows where Luna's missing bugs come from.In Sentry, Discourse and Grafana, Luna came within two verified bugs of Astra. Cal.com had a wider gap, 21 to 30. Keycloak had the widest: Luna found 6 verified bugs to Astra's 14, and only 50% of its Keycloak findings held up, against 93% for Astra.Keycloak is an identity and access management server, and most of its benchmark PRs change authentication and permission logic. The bug-class split points the same way.We labeled every verified bug by root cause. GPT-5.6 Sol, which isn't one of the two models compared here, labeled all 143 bugs in one pass against written definitions. The labels are committed alongside the benchmark data so anyone can check them.On data and logic bugs, the largest group, Luna found 39 to Astra's 47. On concurrency it found 10 to 13. On security, Luna found 9 of 24 and Astra found 19.Two of the Keycloak bugs Astra caught and Luna didn't:Federated recovery codes were never marked as used, so a recovery code could be used more than once.A global view permission overrode denials set on individual clients.Neither looks wrong on any single line. You only see them by working out what the permission model allows after the change.What Luna catches that Astra missesLuna also found bugs Astra missed. Of the 143 verified bugs, 44 were found by both models, 48 only by Astra, and 25 only by Luna.Of the 25 Luna-only bugs, 16 are data and logic bugs and 4 are concurrency bugs. In Discourse, repeating an unsubscribe request kept lowering a user's notification level. In Sentry, a concurrency bug replaced unhealthy worker threads without stopping the old ones.Running both models on every pull request would have found 117 of the 143 verified bugs (82%) for $5.86 in total. That is Luna's $0.20 on top of Astra's $5.66, for 25 more verified bugs.What readers asked us to checkDid the models just remember the fixes?One reader pointed out that these repositories are public, and the fixes for the benchmark bugs may sit in their history. A model trained after those fixes landed could be recalling a patch it has already seen. The suggested test was to split the pull requests by date and see whether the ranking holds on changes made after each model's training cutoff.We can't run that split on this benchmark. We pulled the commit date behind every PR, and they range from 2013 to July 25, 2025. 20 are from 2025, and none are recent enough to fall after either model's cutoff. The post-cutoff group would be empty.The risk is smaller than it sounds, because the defects were added to these PRs for the benchmark on purpose, so the exact bug in each diff is not a commit a model could have trained on. The surrounding code is old and public, though, and a model that knows what the correct version looks like has an advantage. Testing that properly needs pull requests newer than the models, and this benchmark can't provide them.Do the models find the same bugs twice?Another reader asked us to rerun some PRs with identical settings. We picked two PRs per codebase and ran each model two more times.From its first run, Astra had 15 verified bugs on those ten PRs. 10 came back in both repeats and 14 in at least one. Luna also had 15. 7 came back in both repeats and 12 in at least one.The sample is small, so treat these as rough. A model that finds a bug on one run can miss it on the next, which applies to every single-run number in this post, and Luna did it more often than Astra.What about bugs nobody flagged?The third request was to track false negatives, meaning real bugs every model missed. Measuring that needs a complete list of the bugs in each PR, which the benchmark doesn't publish.We can give a lower bound. 26 verified bugs were missed by both Luna and Astra and caught only by Sol or the Entelligence reviewer. Two of them are the Discourse security bugs from our last post: a postMessage origin check that used a substring match, and a remote fetch that followed redirects past a host allowlist. The true number of missed bugs is higher, because bugs no reviewer flagged never enter the pool.Limits of this comparisonApart from the ten repeated PRs, each model reviewed each PR once, and the repeat runs show that results move between runs.Astra is both a contestant and one of the two judges. Requiring Sol to agree reduces the bias without removing it.Every PR predates both models' training cutoffs, so the date split readers asked for isn't possible here.Both models saw the diff and nothing else. They had no repository history, call graph, or production data.Verified counts are a floor on the bugs present, and the benchmark has no complete bug list to measure against.What a diff doesn't tell the modelOn this benchmark, a cheap model did well on most changes and badly on authentication and permission code. A diff alone doesn't tell [2] タイトル: A Beginning for Mathematics URL: https://proofsandprompts.com/2026/09/14/a-beginning-for-mathematics/ 本文: Sep 14, 2026 A beginning for mathematics Daniel Litt, professor at the University of Toronto Three years ago, AI systems could not reliably add two numbers. A year ago, internal models at OpenAI and DeepMind received the equivalent of a gold-medal score on the IMO. Now, these systems are autonomously resolving major open questions. It’s hard to imagine this trend continuing for another year, but I expect it will. It is clear that this will require a radical rethinking of our profession. A few weeks ago, I gave a talk titled The End of Mathematics. If you only read the title1, you might guess that this talk was about how, soon, AI will “solve” math. That’s not what it was about. The talk instead laid out a gloomy vision of the future, in which, despite the possibility of AI systems that are robustly superhuman at mathematics, the design of our institutions causes human understanding of mathematics, and possibly even mathematical progress in the abstract, to stall. I think we will avoid this future, but I also think it is plausibly the default if academic mathematics does not adapt. Despite my relative enthusiasm for the use of AI to do mathematics, I share this view with many of its detractors. Here I want to lay out, instead, a positive vision of the future of mathematics, and the human practice of mathematics. I claim we can deepen human understanding even as the production of interesting mathematics becomes less dependent on it. This essay will take as a premise that AI systems that are robustly superhuman at most or all aspects of mathematics will be here soon. But the concrete changes to our institutions I propose only require accepting the weaker premise that the production of mathematical text is becoming increasingly disconnected from mathematical understanding. What are we even trying to do here? I think it has now become clear that there is no consensus in the mathematical community as to what our goals are. Some of us want to solve problems; some of us think of mathematics as play or as poetry. For some: “Wir müssen wissen – ext{ extendash} wir werden wissen.”2 Some of us think we are penetrating the mysteries of the platonic realm. Some of us think the goal is to embody love of and understanding of mathematics,3 and to transmit that love and understanding to the next generation. My personal, if self-referential, answers are: We’re trying to produce and understand high quality mathematics. We’re trying to produce high quality mathematicians. These goals should be construed broadly. What high quality mathematics consists of has changed quite dramatically over time; we come to its definition as a community. We are not just training PhD students to do research in mathematics. A substantial part of our job, though perhaps an underemphasized one, is to educate the general public about high quality mathematics and mathematical thinking.4 Whatever our goals are, we’ve operationalized them primarily through proving theorems. Almost all papers or PhD theses have a main theorem, and ostensibly a proof of it. But it should be clear that the goal of mathematics is not to prove theorems; if it was, it would be trivial to automate. A computer or monkey could easily start at the axioms of ZFC and iteratively apply deduction rules to them, with no attention whatsoever paid to their meaning. It has had particular significance when a theorem resolves an open problem, especially one that has resisted substantial effort. Again this is easily automated; our computer or monkey can simply conjecture all mathematical propositions in alphabetical order. The general attitude of our community towards a technology that can prove theorems and solve open problems suggests that these operationalizations of our values are at best incomplete. The prospect of automating mathematics by enumerating all conjectures, and all proofs of ZFC, is probably not so disturbing to you. But let us for a moment assume the computer or monkey is very smart; perhaps it understands the results it is proving, and writes beautiful expositions thereof. Perhaps it has a good sense of what we find interesting, and is primarily focusing on those questions. Perhaps it has, in the course of enumerating theorems of ZFC, answered many of our most pressing open questions, and is asking many more fundamental open questions. Is there still a need for human mathematicians? I think so. This machine might produce answers we value, but it would not, in itself, produce human understanding of those answers. In fact I think we are at the beginning of an incredible, wonderful explosion of mathematics, and if we value human understanding, there will be more need for human mathematicians than ever before. But the profession will have to change. In the course of this change, we will have to decide what to hold on to and what to throw away. Some things I would like to preserve: learning seminars; serendipitous conversations that spark an idea; students knocking on a professor’s door to chat about math. A robust community learning exciting new mathematics. Thousands of people that, together, slowly start to resolve their confusion. I worry that much of what has been written on this topic, including some of my own past writing, focuses too much on trying to preserve the precise shape of the institutions of academic mathematics, rather than our values. How can we preserve the journal and peer review system?5 How can we protect the arXiv? How can we keep our role as gatekeepers? If you have internalized the fact that existing AI systems can produce relatively high quality results for the marginal cost of a few dollars, the idea that any semblance of the current equilibrium can survive what’s coming is absurd. As we try to find a new equilibrium, we could try to chase the edge of model capabilities. Right now AI systems arguably underperform us at theory-building, asking questions, exposition, … …so we could prioritize and reward those skills. I think this is unwise: compare the speed at which the academy adapts to the speed at which model capabilities improve. We need to consider the endgame. If the models remain incapable in some domain, we can adjust later. Before I propose some relatively concrete steps we can take, let me remark on what we’re trying to protect mathematics from. There is a lot of anger at AI labs, and certain individuals at those labs. But whatever our judgment of the labs, we need a plan that does not depend on AI capabilities disappearing. The basic issue is not the labs’ behavior, ethical or not.6 It’s the technology itself. I think there is some belief that the labs will “move on” from math next year, be nationalized or broken up, or that a financial bubble will pop, somehow returning things to normal, or… But there is no way our institutions can survive unchanged when anyone with a laptop and a few hundred dollars can generate what would have been an Annals paper last year. AI does not care if you are anti-AI. Producing high-quality mathematicians The most urgent question our profession needs to answer right now is: what should our students be doing? It’s now possible to produce a PhD thesis one hasn’t even read; in terms of demonstrating understanding, mathematical text is worth the paper it is printed on.7 The text no longer reliably conveys a signal about the person who produced it. In my view we should welcome interesting mathematical results regardless of provenance. But our institutions have historically relied on the same signal to indicate both mathematical progress and mathematical expertise. These now must be distinguished. I propose the following reconceptualization of the goal of a mathematics PhD: to become a world expert on some interesting, deep topic, and to be able to convey that interest and understanding to others. Part of operationalizing this might be a thesis, but the degree would be awarded primarily on the basis of a rigorous defense, in which the student ex [3] タイトル: Oracle's Cold 6AM Layoff Emails Hit Staff Amid New Wave of Cuts URL: https://www.techtimes.co.uk/oracle-new-layoffs-restructuring-costs-2-8-billion-1808676 本文: Police Used Flock Camera Footage to Jail Innocent Woman for 13 Days: 'I Thought My Life Was Over' Oracle's Cold 6 AM Layoff Emails Hit Staff Amid New Wave of Cuts: 'Today Is Your Last Working Day' Samsung Says Dua Lipa Has No Case Over $15M Image Lawsuit: 'Nothing Suggesting' She Endorsed TVs Paul Walker's Brother Won't Return in 'Fast Forever': Brian O'Conner Could Still Come Back OpenAI Reveals There Was a Second Rogue AI Incident, Even Before Hugging Face: 'More' May Be Out There