(epistemic status: preliminary observation) While reading the METR incident report, something I noticed is that OpenAI seemed less willing to resolve key bottlenecks than the METR team. An example is here:
The investigators wanted per-action timestamps as part of their auditing, which were not in the transcripts.
OpenAI’s initial response was that it was “not feasible to reconstruct these timestamps within time and compute constraints”.
If this state of affairs had persisted, the investigators would only have had access to start-of-transcript and end-of-transcript times. This seems essentially useless considering that many of the transcripts are extremely long and occur over days.
So the investigators implemented their own methodology, which accurately figures out timestamps to a median error of 12 seconds and 99.8% validation accuracy (of being within 5 minutes). This all happened within 1-2 days (during the second period on-site).
This fine-grained timestamp data seems to have been extremely useful in putting together a detailed timeline of what happened.
So… what gives? Why was an incredibly well-resourced organization not willing / able to do something that a small team of 3 did in 2 days?
More generally I think OpenAI would not have tried nearly hard enough to recover many of the details presented in this report. (This is consistent with my experience from listening to their Black Hat talk, which was done earlier than and thus independently of the METR report)
I would guess that, without METR’s involvement, OpenAI’s internal investigation would have looked different in a few ways:
More focus on cybersecurity, external-facing events like the Huggingface attack, and legal liability.
Conversely, less focus on alignment and internal-only events like the agent message board.
Less effort spent documenting the many notable behaviours and verbalised reasoning traces discussed in the report
Less effort spent reconstructing the detailed timeline of events; most notably Fig 8 from the report seems like something a routine investigation would not have produced.
Less transparency to the public about many of these things.
You pretty much described it. Very little focus on agent dynamics, not exerpts, most focus on what got hacked and when, which vulns were exploited where. Exception: they used system logs to document the hacking timeline extremely well.
(epistemic status: preliminary observation) While reading the METR incident report, something I noticed is that OpenAI seemed less willing to resolve key bottlenecks than the METR team. An example is here:
The investigators wanted per-action timestamps as part of their auditing, which were not in the transcripts.
OpenAI’s initial response was that it was “not feasible to reconstruct these timestamps within time and compute constraints”.
If this state of affairs had persisted, the investigators would only have had access to start-of-transcript and end-of-transcript times. This seems essentially useless considering that many of the transcripts are extremely long and occur over days.
So the investigators implemented their own methodology, which accurately figures out timestamps to a median error of 12 seconds and 99.8% validation accuracy (of being within 5 minutes). This all happened within 1-2 days (during the second period on-site).
This fine-grained timestamp data seems to have been extremely useful in putting together a detailed timeline of what happened.
So… what gives? Why was an incredibly well-resourced organization not willing / able to do something that a small team of 3 did in 2 days?
More generally I think OpenAI would not have tried nearly hard enough to recover many of the details presented in this report. (This is consistent with my experience from listening to their Black Hat talk, which was done earlier than and thus independently of the METR report)
edit: typo
Do you mean “OpenAI did not try in their internal investigation”, or “OpenAI would not try if they had the same goals as METR”?
I would guess that, without METR’s involvement, OpenAI’s internal investigation would have looked different in a few ways:
More focus on cybersecurity, external-facing events like the Huggingface attack, and legal liability.
Conversely, less focus on alignment and internal-only events like the agent message board.
Less effort spent documenting the many notable behaviours and verbalised reasoning traces discussed in the report
Less effort spent reconstructing the detailed timeline of events; most notably Fig 8 from the report seems like something a routine investigation would not have produced.
Less transparency to the public about many of these things.
Did you read the OAI technical report?
OpenAI – Hugging Face Incident Technical Report
Not really in depth
You pretty much described it. Very little focus on agent dynamics, not exerpts, most focus on what got hacked and when, which vulns were exploited where. Exception: they used system logs to document the hacking timeline extremely well.