METR
An organization in the news record
| Kind | An organization · graph entity E2 · Wikidata Q135185153 |
|---|---|
| Wikipedia's sources | 0 of the 0 references dated 2026-09-01 to 2026-11-21 in METR · 3 more from the year before the record |
| Views | 72 quoted, from 5 source domains |
| Dates | 2026-09-23 to 2026-11-20 · 6 weeks |
| Arguments | 10 · AI accelerating AI R&D, Near-term AI R&D automation, Alignment eval validity and eval awareness, Scope of AI developer risk reports, AI accelerating AI and science, AI incidents and loss of control … |
| Reported by | Dwarkesh Patel, Latent Space, METR, Transformer, Zvi Mowshowitz |
| Profile | a readable profile, written from these views and checked sentence by sentence |
Only what the organization says as itself. Views of its staff, teams or founders, named or not, are not counted here.
Where they stand
By argument, from their own words. ✓ agree · ✗ disagree · ~ no clear side · · side unclear (the two readings differ).
AI incidents and loss of control ~
- ~
It recommends treating all transcripts and actions as untrusted input.
AI accelerating AI R&D ✓ ~
- ✓
we expect AI’s augmentation of R&D to continue to grow, and we cannot rule out a substantial further acceleration in the pace of capabilities progress over the next 6 months.
- ✓
We guess that a majority (perhaps ≈80%) of the current (October 2026) acceleration is due to AI.
- ✓
we believe there is a significant risk that further capabilities progress could lead to even faster acceleration over the coming year.
- ✓
We did not see evidence for significant bottlenecks in algorithmic progress, training compute, experiment compute, or data that will take effect in the next 12 months. However, our understanding of the status of different bottlenecks changed several times
- ~
We received convincing evidence that the acceleration was not due to an acceleration in the scale of training compute.
- ~
the public has to rely on METR’s judgment much more than would be ideal, given the weight of the questions answered here.
- ~
We consider this to be only very loose evidence, as the responses reflect low-effort consideration.
- ~
we think it is extremely unlikely Anthropic was actively trying to mislead us.
- ✓ Quote not reprinted: this page already repeats as much of this article as we allow.
- ~ Quote not reprinted: this page already repeats as much of this article as we allow.
Near-term AI R&D automation ✓ ~ ·
- ✓
We think that current public evidence is insufficient to rule out drastic AI R&D acceleration in the next couple of years.
- ✓
It may in fact be possible to drastically accelerate AI R&D via easy-to-verify tasks alone.
- ·
The most straightforward explanation for this is that models have poor judgment about what makes a good task.
- ✓
We find that current public frontier models have extremely strong capabilities on optimization tasks, often outperforming skilled humans.
- ·
We are increasingly concerned that our easy-to-run, containerized evals could overestimate real-world capabilities, due to similar RL environments becoming common in training.
- ·
we would consider weakness in ‘judgment’ and related capabilities to be reasonable candidates for what most delays drastic AI R&D acceleration.
- ~
One possibility that we find plausible is that doubling times may tend to be shorter on less diverse sets of tasks.
- ✓ Quote not reprinted: this page already repeats as much of this article as we allow.
- · Quote not reprinted: this page already repeats as much of this article as we allow.
Alignment eval validity and eval awareness ✓ ~ ·
- ✓
alignment evals built the usual way may be easy for current models to tell apart from ordinary tasks
- ✓
The presence of tells (like this “toyness”) could lead alignment evaluations to over-estimate a model’s alignment
- ·
the latest Anthropic models seem to be better than OpenAI models
- ~
We think it will be easier to get rid of unintentional environment-level tells than intrinsic task-level tells, although both might be possible.
- ·
it can bias the agent’s answer away from what it would actually believe if it had not been asked the question
- ~
This variance seemed to be because the agents were under-elicited on some tasks
- ✓
Understanding alignment tells can help us make our alignment evals more on-distribution relative to deployment or capability evals.
Scope of AI developer risk reports ·
- ·
We think that undermining safety-relevant decision-making should be explicitly within the scope of risk pathway 2 (“Targeted undermining of safety-relevant R&D”) or a separate risk pathway.
- ·
we think risk reports should explicitly reason about whether models that are still in training, but not yet internally deployed, should be in scope.
AI accelerating AI and science ✓
- ✓
~1.5X overall acceleration in capabilities due to AI (i.e. 1.5 years in 1 year), with perhaps 30% chance of 2X acceleration.
Containing AI with classical security ~
- ~
we believe that future monitoring infrastructure should be able to withstand deliberate, strategic, and sustained attacks from highly capable internal agents at scale.
Automated R&D catastrophic risk ✓
- ✓
we broadly agree that direct catastrophic risks due to automated R&D from models on July 15th were low.
benchmark saturation ✗
- ✗ Quote not reprinted: this page already repeats as much of this article as we allow.
reward hacking effort in evaluation development ·
- · Quote not reprinted: this page already repeats as much of this article as we allow.
Said differently
None in the record.
History
- · 23 views
- · on reward hacking effort in evaluation development: Quote not reprinted: this page already repeats as much of this article as we allow.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- · on Near-term AI R&D automation: Quote not reprinted: this page already repeats as much of this article as we allow.
- ✓ on Near-term AI R&D automation: Quote not reprinted: this page already repeats as much of this article as we allow.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- ~ on Near-term AI R&D automation:
One possibility that we find plausible is that doubling times may tend to be shorter on less diverse sets of tasks.
- · on Near-term AI R&D automation:
we would consider weakness in ‘judgment’ and related capabilities to be reasonable candidates for what most delays drastic AI R&D acceleration.
- · on Near-term AI R&D automation:
We are increasingly concerned that our easy-to-run, containerized evals could overestimate real-world capabilities, due to similar RL environments becoming common in training.
- ✓ on Near-term AI R&D automation:
We find that current public frontier models have extremely strong capabilities on optimization tasks, often outperforming skilled humans.
- · on Near-term AI R&D automation:
The most straightforward explanation for this is that models have poor judgment about what makes a good task.
- ✓ on Near-term AI R&D automation:
It may in fact be possible to drastically accelerate AI R&D via easy-to-verify tasks alone.
- ✓ on Near-term AI R&D automation:
We think that current public evidence is insufficient to rule out drastic AI R&D acceleration in the next couple of years.
Most of the improvement between Opus 4.6 and Opus 4.8 comes from recognizing alignment evals
- ✓ on Alignment eval validity and eval awareness:
Understanding alignment tells can help us make our alignment evals more on-distribution relative to deployment or capability evals.
- ~ on Alignment eval validity and eval awareness:
This variance seemed to be because the agents were under-elicited on some tasks
- · on Alignment eval validity and eval awareness:
it can bias the agent’s answer away from what it would actually believe if it had not been asked the question
- ~ on Alignment eval validity and eval awareness:
We think it will be easier to get rid of unintentional environment-level tells than intrinsic task-level tells, although both might be possible.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- · on Alignment eval validity and eval awareness:
the latest Anthropic models seem to be better than OpenAI models
- Quote not reprinted: this page already repeats as much of this article as we allow.
- ✓ on Alignment eval validity and eval awareness:
The presence of tells (like this “toyness”) could lead alignment evaluations to over-estimate a model’s alignment
- ✓ on Alignment eval validity and eval awareness:
alignment evals built the usual way may be easy for current models to tell apart from ordinary tasks
- · 3 views
- Quote not reprinted: this page already repeats as much of this article as we allow.
We are uncertain, but think it’s unlikely that [Claude Mythos 5.2] can fully automate the AI R&D process at Anthropic.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- · 21 views
- ~ on AI accelerating AI R&D: Quote not reprinted: this page already repeats as much of this article as we allow.
- ✗ on benchmark saturation: Quote not reprinted: this page already repeats as much of this article as we allow.
- ✓ on AI accelerating AI R&D: Quote not reprinted: this page already repeats as much of this article as we allow.
- ~ on AI accelerating AI R&D:
we think it is extremely unlikely Anthropic was actively trying to mislead us.
- ~ on AI accelerating AI R&D:
We consider this to be only very loose evidence, as the responses reflect low-effort consideration.
- ~ on AI accelerating AI R&D:
the public has to rely on METR’s judgment much more than would be ideal, given the weight of the questions answered here.
- ~ on AI accelerating AI R&D:
We received convincing evidence that the acceleration was not due to an acceleration in the scale of training compute.
- ✓ on AI accelerating AI R&D:
We did not see evidence for significant bottlenecks in algorithmic progress, training compute, experiment compute, or data that will take effect in the next 12 months. However, our understanding of the status of different bottlenecks changed several times
- ✓ on AI accelerating AI R&D:
we believe there is a significant risk that further capabilities progress could lead to even faster acceleration over the coming year.
- ✓ on AI accelerating AI R&D:
We guess that a majority (perhaps ≈80%) of the current (October 2026) acceleration is due to AI.
- Quote not reprinted: this page already repeats as much of this article as we allow.
The public risk report states that Model 2’s score has large error bars, but does not state the qualification in the private risk report that Anthropic expects higher scores from later snapshots of the same base model.
The lack of such benchmark evidence makes us more worried that alignment, tooling, oversight or elicitation issues are holding back automation
- ✓ on AI accelerating AI R&D:
we expect AI’s augmentation of R&D to continue to grow, and we cannot rule out a substantial further acceleration in the pace of capabilities progress over the next 6 months.
given the acceleration in the leading indicators mentioned above, a sustained acceleration seems more parsimonious.
- Quote not reprinted: this page already repeats as much of this article as we allow.
A better description would seem to be “around 2.”
- ✓ on Automated R&D catastrophic risk:
we broadly agree that direct catastrophic risks due to automated R&D from models on July 15th were low.
we believe the public report overall understates the level of acceleration that was occurring
- Quote not reprinted: this page already repeats as much of this article as we allow.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- · 23 views
while risk of catastrophic harm from misalignment i s low for now, this is mostly because the models aren’t smart enough — not because they’re sufficiently aligned or under control.
The exercise benefited substantially from the depth of access described above, and from significant logistics and technical support from OpenAI staff.
Periodically perform control evaluations, search for exploitable bugs in the monitoring system, and develop methodology to evaluate strategic collusion.
actions agents take should not be able to influence whether or when the monitoring system runs.
Running unmonitored requests should require explicit approval to use a specific set of organizations/API keys, which aren’t allowed to be used for agentic inference.
Agents should be monitored every N messages/minutes instead of waiting for them to end their turns.
We estimated that not monitoring compaction events likely led to on the order of 10% of eligible internal Codex traffic not being monitored between May and July 2026
- Quote not reprinted: this page already repeats as much of this article as we allow.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- ~ on Containing AI with classical security:
we believe that future monitoring infrastructure should be able to withstand deliberate, strategic, and sustained attacks from highly capable internal agents at scale.
- · on Scope of AI developer risk reports:
we think risk reports should explicitly reason about whether models that are still in training, but not yet internally deployed, should be in scope.
we find no evidence that Mythos 5 was deliberately sandbagging its critiques in order to mislead Anthropic about the level of catastrophic risks (and believe that this is a priori unlikely)
we think it’s plausible that Claude’s sycophancy degrades Anthropic’s epistemics in mundane (i.e., not catastrophic) situations.
Although Anthropic acknowledges some of these limitations, we think they don’t do so strongly enough.
we think it is plausible that covered models can effectively obfuscate some types of severely misaligned actions through CoT rationalization when monitored by systems that read their reasoning.
- · on Scope of AI developer risk reports:
We think that undermining safety-relevant decision-making should be explicitly within the scope of risk pathway 2 (“Targeted undermining of safety-relevant R&D”) or a separate risk pathway.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- Quote not reprinted: this page already repeats as much of this article as we allow.
- · 1 views
- ~ on AI incidents and loss of control:
It recommends treating all transcripts and actions as untrusted input.
- ~ on AI incidents and loss of control:
- · 1 views
- ✓ on AI accelerating AI and science:
~1.5X overall acceleration in capabilities due to AI (i.e. 1.5 years in 1 year), with perhaps 30% chance of 2X acceleration.
- ✓ on AI accelerating AI and science:
Built 2026-11-21 06:50 UTC from the news knowledge
graph by code. Every quote was checked against the source's own words; which side a quote takes is a model's reading,
made twice.