The most prestigious law school admissions discussion board in the world.
|
 |
 
|
Mythos is actually a step change in AI capabilities
Poast new message in this thread
Date: June 9th, 2026 9:52 PM Author: zippy provocative business firm https://www.oneusefulthing.org/p/what-it-feels-like-to-work-with-mythos
https://x.com/victortaelin/status/2064448425936994742?s=46
https://metr.org/time-horizons/
https://www.anthropic.com/news/claude-fable-5-mythos-5
Meanwhile it’s showing large improvements on the sorts of software engineering tasks necessary for recursive self-improvement. EPAH crying, losing hope. |
Date: June 9th, 2026 9:55 PM Author: haunting station that step change in question?
it's doing the exact same stuff it did before (language tasks), but slightly better
Wow. A Step Change Was Performed. This. Changes. Everything. |
 |
Date: June 9th, 2026 9:59 PM Author: zippy provocative business firm Language tasks includes writing code for self-improving AGI. This is not comforting (http://www.autoadmit.com/thread.php?thread_id=5872645&forum_id=2",#49926695) |
|
Date: June 9th, 2026 10:06 PM Author: haunting station no, they include writing code for self-improving language machines (LLMs)
seems like a safe bet that they will keep getting better at language tasks. wow, more computer stuff on the computers. Game. Changer. |
|
Date: June 9th, 2026 9:56 PM Author: dun point yeah it beat Pokemon FireRed without a GameFAQs walkthrough which is something no brainrotten zoomer could ever achieve (http://www.autoadmit.com/thread.php?thread_id=5872645&forum_id=2",#49926687) |
|
Date: June 9th, 2026 10:02 PM Author: Fear-inspiring abode tp (http://www.autoadmit.com/thread.php?thread_id=5872645&forum_id=2",#49926705) |
|
Date: June 9th, 2026 10:02 PM Author: Fear-inspiring abode *pumps shotgun* (http://www.autoadmit.com/thread.php?thread_id=5872645&forum_id=2",#49926703) |
|
Date: June 9th, 2026 10:05 PM Author: motley home party of the first part it instructed me to write a Paper on AI Alignment and cold-email some executives (http://www.autoadmit.com/thread.php?thread_id=5872645&forum_id=2",#49926708) |
|
Date: June 9th, 2026 10:05 PM Author: haunting station (http://www.autoadmit.com/thread.php?thread_id=5872645&forum_id=2",#49926711) |
|
Date: June 9th, 2026 10:06 PM Author: chestnut swashbuckling forum Just for coders or will it give great legal answers with no hallucinations? (http://www.autoadmit.com/thread.php?thread_id=5872645&forum_id=2",#49926712) |
|
Date: June 9th, 2026 11:46 PM Author: smoky jet-lagged box office just for coders because its the only thing SV needs it to be good at and then everything else is just rhlf trained by braindead lib "experts" and safety alignment people. I'm sure it's good at math too. (http://www.autoadmit.com/thread.php?thread_id=5872645&forum_id=2",#49926899) |
|
Date: June 9th, 2026 11:55 PM Author: zippy provocative business firm 8.10 USAMO 2026 The USA Mathematical Olympiad (USAMO) is a six-problem, two-day proof-based competition for high school students. It is the next step of the math olympiad track in the US after the AIME, which was a popular AI benchmark last year but is now saturated. The 2026 USAMO took place on March 21–22, 2026, after almost all of Mythos’s training data was collected, and we are confident that there was no contamination. Because USAMO solutions are proofs rather than short answers, grading can be challenging and subjective. We follow the MathArena 41 grading methodology, where each proof is rewritten by a neutral model (Gemini 3.1 Pro) and judged by a panel of 3 frontier models (we used Gemini 3.1 Pro, Claude Opus 4.6, and Claude Mythos Preview) according to defined rubrics. The final score is the minimum given by any judge. Mythos 5 scored 99.8% at medium, high, and xhigh reasoning effort, and 98.3% at low effort, averaging over 10 attempts per problem. Across all 240 attempts, the only proof that more than one judge scored below full marks was a low-effort attempt on Problem 6, where the model itself declined to claim a complete solution and proved a restricted subcase instead. Average token usage per attempt ranged from roughly 42K at low effort to 100K at xhigh. Under similar settings, Opus 4.8 scored 96.7% and Opus 4.7 scored 69.3%
-----------
an LLM is definitely going to 100% the IMO this year. |
|
Date: June 10th, 2026 12:00 AM Author: smoky jet-lagged box office Yeah it's also way better than that though since it can solve Erdos problems and shit. Olympiad problems haven't been hard for AI for a while, it just gets penalized for not writing the "proofs" exactly like a precocious ape child would (http://www.autoadmit.com/thread.php?thread_id=5872645&forum_id=2",#49926921) |