Загрузка видео...

Не удалось загрузить видео

На главную

you added a second model to check the first one it never opened the repository what you handed it was the summary, so what it reviewed was a paragraph, and the paragraph was written by something fluent, which is why it reads as plausible two models agreeing about a...

28,822 просмотров • 7 дней назад •via X (Twitter)

Комментарии: 0

Нет доступных комментариев

Здесь появятся комментарии из оригинального поста

Похожие видео

#WATCH | On UGC Regulations 2026, Dr Vikas Divyakirti, founder and MD of Drishti IAS, says, "...No one is even talking about what happened to one or two children from the Northeast in Delhi. While the preamble of the Equity Committee even mentions the word 'race,' and also mentions 'place of birth, ' someone's place of birth is Nagaland, or let's say Mizoram, every person in Delhi is using strange language towards them, and that's not even addressed in the Equity Committee. In the Equity Committee, there is no representation for the people of different races, and no one is addressing it...The Equity Committee lacks representation of linguistic diversity. There is no representation at all. Why is it only about the OBC versus the General Category, and not about other categories?... You initially took six bases: religion, race, caste, gender, place of birth, and disability...When discrimination is mentioned, a whole paragraph is written, and in the discrimination paragraph, it clearly states religion, race, caste, gender, place of birth, disability, and any of them – there are six. Why did you need to write a separate paragraph on caste-based discrimination? When discrimination was already mentioned, why write a separate paragraph on caste-based discrimination? And if you did, why didn't you write a paragraph on religion, on race, on gender? All the categories. You should have written about all of them. Why are you segregating only one basis as a special basis in the definitions? I think those who drafted these rules haven't read them very carefully...The Supreme Court date was approaching, so they passed it in a hurry..." Watch full interview at -

ANI

18,784 просмотров • 7 месяцев назад

your eval measured one system. production ran another. they start from the same idea and end somewhere else, which is why the number was right and the outcome was not. the code is the same until someone rewrites it for prod, because the eval lived in a notebook and prod is a service. then the data diverges. clean fixtures in the test, late and incomplete data in the run. then execution diverges. the eval assumed every call lands. production has retries, timeouts and rate limits. by the end the two runs share a name and nothing else. > why nobody catches it both sides are internally correct. the eval measured what it ran. production ran what it was given. neither one is wrong, and neither one is responsible for checking they were the same artifact. so the gap widens quietly. someone patches prod without touching the eval. someone adds a fixture without touching prod. no test anywhere fails. > what to actually do ship one artifact. the code that was measured is the code that runs. model the friction. whatever the real environment charges you, charge it in the eval too. easiest way to see the gap on your own idea, without writing the harness: (only first 100 registrated) it runs the tested code live, unchanged. that is the whole point. and treat the eval as a prior, not a result. it tells you what to expect. only the live run tells you what happened. your eval was accurate. it was accurate about something you never shipped.

Hanako

45,239 просмотров • 20 дней назад

your agent reviewing its own work is not a check. it is a second opinion from the same source. this is the most common gap in agent systems and it hides in plain sight, because the step exists. there is a review. it just cannot do the thing you think it does. here is the mechanism. the model produced an output from a context. you then ask the same model, holding the same context, whether that output is correct. it answers fluently, because that is what it does. and the answer is drawn from the same distribution that produced the thing being judged. same weights, same window, same blind spots. if the reason the output is wrong is something the model does not know, the review does not know it either. if the reason is something the context does not contain, the review has the same context. the failure mode and the detector share a cause. > why it feels like it works because most of the time the output is fine, and the review says fine. agreement is not evidence of detection. a reviewer that says pass on everything agrees with reality most of the time too. what you actually want to measure is what happens on the cases that are wrong. that is the only place a check earns its name, and it is exactly the place where a self-review is weakest. there is research on this. Huang and colleagues at DeepMind showed at ICLR 2024 that intrinsic self-correction, revising without external grounding, does not reliably help and often makes things worse. > what to actually do move the check outside the model. a test that runs, a schema that validates, a file that exists or does not, an exit code from something you did not write. these are not smarter than the model. they are just not correlated with it, and that is the entire value. when the judgement genuinely needs a model, at minimum use a different family. same family means shared blind spots, and frontier judges measurably inflate scores for outputs that look like their own. and split the work by kind. anything objectively checkable goes to code. only the genuinely semantic calls go to a judge, and those get a rubric written as one line. a review inside the loop tells you the model is confident. a check outside it tells you whether the work is done. save this - then read the eval setup below

Hanako

14,325 просмотров • 24 дней назад