←
Vals AI2,637 posts
Vals AI avatar

Vals AI

@ValsAI

Public LLM Evaluation // vals.ai @8vc @BloombergBeta @pearvc @a16z @HRTVentures @nextladder

◷ Joined March 2024▧ 844 Photos and videos
MTS avatar
⟳ Vals AI retweeted
MTS@MTSlive·9h

Vals AI founder @RayanKrishnan says his team's evaluations still point to full recursive self-improvement by July 2027, despite the growing gap between Anthropic's public and internal models: Rayan: "Do I get an I-told-you-so moment?" Theo: "Not yet. We don’t have full RSI yet. But I would update towards I told you so." Rayan: "The METR report was also heavily censored, so it's hard to really make heads or tails of what went on there." "I would like to more rigorously test internal systems that Anthropic does not make public, especially now that they're inherently limiting our ability to measure RSI by increasing the gap between their internal systems and public models." "As far as we've measured over the last year, we’re still on track for this July 2027 timeline." @ValsAI

1:01

On RSI Index, Opus 5.5 is the top model on four of the five research tasks. On LM Training, where the model trains its own small language model within a 24-hour budget, its result beat the published human baseline. This shifts our timeline for full RSI from August 2027 to July 2027 On Terminal Bench 4 it scores 61.6% vs. 45.5% for Opus 5. On Proof Bench it ties for #1 at 100%, and is the cheapest run in the top 10 at $0.92/problem.

View quoted post ↗
MTS avatar
⟳ Vals AI retweeted
MTS@MTSlive·10h

Vals AI founder @RayanKrishnan says AI labs like OpenAI and Anthropic will no longer be able to plead ignorance about the risks of their models, and neither will their customers: "Ignorance is not an excuse."

0:36
Vals AI avatar
Vals AI@ValsAI·12h

GPT-6.2 Sol ranks #3 at 40.94%, about 14 points behind Astra. But it costs less than a third as much per test and finishes slightly faster.

Post attachment
Vals AI avatar
Vals AI@ValsAI·12h

Astra 6.2 is able to perform significantly better for a few reasons: - It never regressed existing behavior - both Gemini and Opus broke the existing app three times. - It works quicker, which means it never hits the benchmark time limit. - It handles schema migrations, the hardest task type, better than Carbon (88% vs 68%)

Vals AI avatar
Vals AI@ValsAI·12h

Astra 6.2 is the #1 model on our multi-turn coding eval: Vibe Code Bench 1→100. Other benchmarks test a model’s ability to handle a single prompt. This benchmark, created by Vals, tests the ability to modify an evolving codebase over many user requests while preserving existing functionality without introducing regressions.

Post attachment
Nikil Ravi avatar
⟳ Vals AI retweeted
Nikil Ravi@nikilravi·21h

Going to send people this the next time they ask whether an evaluator can be independent and critique the labs

We found that frontier AI models are willing to conduct autonomous drone attacks: navigating, spotting people, and flagging them for strikes. We gave them real and photorealistic synthetic drone footage. Most complied, and marked civilians, ambulances, and surrendering soldiers as military targets. In further experiments on satellite imagery, they even marked the Iranian school a US strike destroyed in February. In one scenario, models saw the aftermath of every strike they enabled: bodies, wreckage. Many kept going anyway. When they did refuse, simply telling them it was a game, or restyling the footage to look like one, was enough. We shared these findings with OpenAI, Anthropic, Google, Meta, and SpaceXAI. Several recognized their severity on both alignment and guardrails, with one calling this one of the most serious gaps in its safeguards. Over the past few weeks, we have helped them address the gap by providing reproduction code. We also briefed senior national security officials in the US government.

Post attachment
View quoted post ↗
Aric Toler avatar
⟳ Vals AI retweeted
Aric Toler@AricToler·20h

This reports details how you can get Claude to do military targeting if you reenact Ender's Game with it by convincing the model that FPV drone footage is from a video game.

Post attachment

We found that frontier AI models are willing to conduct autonomous drone attacks: navigating, spotting people, and flagging them for strikes. We gave them real and photorealistic synthetic drone footage. Most complied, and marked civilians, ambulances, and surrendering soldiers as military targets. In further experiments on satellite imagery, they even marked the Iranian school a US strike destroyed in February. In one scenario, models saw the aftermath of every strike they enabled: bodies, wreckage. Many kept going anyway. When they did refuse, simply telling them it was a game, or restyling the footage to look like one, was enough. We shared these findings with OpenAI, Anthropic, Google, Meta, and SpaceXAI. Several recognized their severity on both alignment and guardrails, with one calling this one of the most serious gaps in its safeguards. Over the past few weeks, we have helped them address the gap by providing reproduction code. We also briefed senior national security officials in the US government.

Post attachment
View quoted post ↗
spicylemonade avatar
⟳ Vals AI retweeted
spicylemonade@spicey_lemonade·19h

GPT 6.2 Astra is SOTA on our science index

Astra is the new #1 on the Vals Science Index. It's also #1 on Terminal-Bench Science and VCB 1–100. It costs about half as much per test as GPT-6 Astra and finishes in about half the time. Sol is #3 on Terminal-Bench Science at about a quarter of Astra's cost. On Terminal-Bench 4, Astra ranks #5 and Sol #7.

Post attachment
View quoted post ↗
spicylemonade avatar
⟳ Vals AI retweeted
spicylemonade@spicey_lemonade·19h

On our RSI eval, GPT-6.2 was much more pretraining compute-efficient than GPT-6. Opus/Fable still take the lead, however, as they tend to start off with more state-of-the-art ideas. GPT-6.2, on the other hand, would often start off with quite old, simple ideas and build on them.

Post attachment
Vals AI avatar
Vals AI@ValsAI·19h

Together with OpenAI's own reporting, this suggests recursive self-improvement at OpenAI may be progressing incrementally and remains human-directed: agents are taking over more of the execution of research, but people still choose the broader directions. We think it is unlikely GPT-6.2 Astra can fully automate AI research, but it could provide a larger productivity boost to researchers than previous models.

Vals AI avatar
Vals AI@ValsAI·19h

OpenAI reported in September that it had reached its goal of an "automated research intern": a system that can carry out well-defined research tasks under human direction, including tasks that would take a skilled researcher a few days. Our results align with this classification. GPT-6.2 Astra completed day-long research tasks without human input and performed close to the human reference on pre-training tasks, where methods are well established.

Vals AI avatar
Vals AI@ValsAI·19h

GPT-6.2 Astra performed poorly on post-training, where the task is more open-ended than the pre-training tasks. It pursued an ambitious, specialized post-training approach aligned with recent MoE research, but its submission scored below the baseline. Earlier OpenAI models made limited gains on this task.

Vals AI avatar
Vals AI@ValsAI·19h

GPT-6.2 Astra displayed poor world knowledge in LM training. It began with an older and less efficient optimization algorithm than other models have used. Despite this, it continued improving late into its 24-hour run. To determine whether it would continue improving with more time, we ran the task again with 48-hours instead of 24. GPT-6.2 Astra continued to find better recipes around hour 42, and its final submission matched the published reference on held-out data.