Embed Notice

HTML Code

<blockquote style="position: relative; padding-left: 55px;"><section><a href="https://mastodon.social/users/shriramk/statuses/116998675100262348">Shriram Krishnamurthi (shriramk@mastodon.social)'s status on Wednesday, 29-Jul-2026 02:10:04 JST</a><a href="https://mastodon.social/@shriramk" title="shriramk@mastodon.social"><img src="https://gnusocial.jp/avatar/79575-48-20221223221604.webp" width="48" height="48" alt="Shriram Krishnamurthi" style="position: absolute; left: 0; top: 0;">Shriram Krishnamurthi</a><div><a href="https://mastodon.social/@shriramk/116998671547046391" rel="in-reply-to">in reply to</a></div></section><article><p>1/ ↵</p></article><footer><a rel="bookmark" href="https://gnusocial.jp/conversation/6583866#notice-12962053">In conversation</a><time datetime="2026-07-29T02:10:04+09:00" title="Wednesday, 29-Jul-2026 02:10:04 JST">about 3 months ago</time> <span>from <span><a href="https://mastodon.social/@shriramk/116998675100262348" rel="external" title="Sent from mastodon.social via ActivityPub">mastodon.social</a></span></span><a href="https://mastodon.social/@shriramk/116998675100262348">permalink</a><h4>Attachments</h4><ol><li><label><a rel="external" href="https://gnusocial.jp/attachment/6295815">&gt; **Historically, CAV was focused on the automatic discovery of code, proofs, and counterexamples. There are two ways AI could make it irrelevant: AI could make our problems too easy, or AI could ma**(a) A lot of tasks are really not that complicated and don't depend a lot on verification, and the training is good enough. It's like asking a PhD student to write a sorting function: you don't worry a lot that it might be done wrong and that you need to deeply verify it.(b) The current success of AI coding is _because_ of verification. We didn't get here through LLMs alone; on their own, they didn't do that great a job beyond basic tasks. We got here because of the agentic loop, where the agents are able to check their work. We use different levels of checking — some of the "verification" is closer to the level of fuzzing (does the program even parse) — but we are steadily moving up the hierarchy, and better generation is tied intimately to better checking. So verification, broadly, is indispensable to current successes.(c) I'd really rather not talk about "proving things right" or even "trust" but rather "confidence". We _gain confidence_, and we do so by many means. And that's what we've always done with software.</a></label><br><a href="https://files.mastodon.social/media_attachments/files/116/998/671/911/373/261/original/fec9e0d9930b0ac8.png" rel="external">https://files.mastodon.social/media_attachments/files/116/998/671/911/373/261/original/fec9e0d9930b0ac8.png</a></li></ol></footer></blockquote>

Corresponding Notice

Embed this notice
Shriram Krishnamurthi (shriramk@mastodon.social)'s status on Wednesday, 29-Jul-2026 02:10:04 JST Shriram Krishnamurthi
in reply to
1/ ↵
In conversationabout 3 months ago from mastodon.socialpermalink
Attachments
1. > **Historically, CAV was focused on the automatic discovery of code, proofs, and counterexamples. There are two ways AI could make it irrelevant: AI could make our problems too easy, or AI could ma** (a) A lot of tasks are really not that complicated and don't depend a lot on verification, and the training is good enough. It's like asking a PhD student to write a sorting function: you don't worry a lot that it might be done wrong and that you need to deeply verify it. (b) The current success of AI coding is _because_ of verification. We didn't get here through LLMs alone; on their own, they didn't do that great a job beyond basic tasks. We got here because of the agentic loop, where the agents are able to check their work. We use different levels of checking — some of the "verification" is closer to the level of fuzzing (does the program even parse) — but we are steadily moving up the hierarchy, and better generation is tied intimately to better checking. So verification, broadly, is indispensable to current successes. (c) I'd really rather not talk about "proving things right" or even "trust" but rather "confidence". We _gain confidence_, and we do so by many means. And that's what we've always done with software.
  https://files.mastodon.social/media_attachments/files/116/998/671/911/373/261/original/fec9e0d9930b0ac8.png

Public

Embed Notice

HTML Code

Corresponding Notice