The original RSP text uses the word commit for the relevant quote. They could have described it differently in the text itself if they weren't sure about meeting that standard in the future.
IMO it's less about the object level badness of the change, which is small potatoes compared to many other recent cases from other leading AI companies, but more about the meta level point that commitments that can be changed aren't worth very much.
Not meaning to imply that Anthropic has dropped ASL-4! Just wanted to call out that this is does represent a change from the Sept. 2023 RSP.
I wrote the article Mikhail referenced and wanted to clarify some things.
The thresholds are specified, but the original commitment says, "We commit to define ASL-4 evaluations before we first train ASL-3 models (i.e. before continuing training beyond when ASL-3 evaluations are triggered). Similarly, we commit to define ASL-5 evaluations before training ASL-4 models, and so forth," and, regarding ASL-4, "Capabilities and warning sign evaluations defined before training ASL-3 models."
The latest RSP says this of CBRN-4 Required Safeguards, "We expect this threshold will require the ASL-4 Deployment and Security Standards. We plan to add more information about what those entail in a future update."
Additionally, AI R&D 4 (confusingly) corresponds to ASL-3 and AI R&D 5 corresponds to ASL-4. This is what the latest RSP says about AI R&D 5 Required Safeguards, "At minimum, the ASL-4 Security Standard (which would protect against model-weight theft by state-level adversaries) is required, although we expect a higher security standard may be required. As with AI R&D-4, we also expect an affirmative case will be required."
You're right that contradicted is too strong a word here, though I think OpenAI's new claim does cast doubt on the earlier reported claim.
I think the fact that investors appear to be fine with this new arrangement is the biggest tell that it's not a very significant change from the original plan. OpenAI's nonprofit mission is to ensure AGI benefits humanity, not that it be the first to build it. Legally, the organization has to show that its charitable purpose isn't changing or that it has a sufficient justification to change it.
There's more context in my past coverage of the restructuring effort.
Sorry, lot on my plate.
You're basically asking how we'd operationalize the claim that either the USG or PRC are "racing toward AGI"? Probably would involve some dramatic action like consolidating large amounts of compute into projects that are either nationalized or contracted to the govt (like this part of AI-2027:
"A Centralized Development Zone (CDZ) is created at the Tianwan Power Plant (the largest nuclear power plant in the world) to house a new mega-datacenter for DeepCent, along with highly secure living and office spaces to which researchers will eventually relocate. Almost 50% of China’s AI-relevant compute is now working for the DeepCent-led collective,38 and over 80% of new chips are directed to the CDZ.39 At this point, the CDZ has the power capacity in place for what would be the largest centralized cluster in the world."
Do you want to suggest specific thresholds or modifications?
I don't think it's fair to say I made a bad prediction here.
Here's the full context of my quote: "The report clocks in at a cool 793 pages with 344 endnotes. Despite this length, there are only a handful of mentions of AGI, and all of them are in the sections recommending that the US race to build it.
In other words, there is no evidence in the report to support Helberg’s claim that "China is racing towards AGI.”
Nonetheless, his quote goes unchallenged into the 300-word Reuters story, which will be read far more than the 800-page document. It has the added gravitas of coming from one of the commissioners behind such a gargantuan report.
I’m not asserting that China is definitively NOT rushing to build AGI. But if there were solid evidence behind Helberg’s claim, why didn’t it make it into the report?"
Here's my tweet mentioning Gwern's comment. It's not clear that DeepSeek falsifies what Gwern said here:
the scientific culture of China is 'mafia' like (Hsu's term, not mine) and focused on legible easily-cited incremental research, and is against making any daring research leaps or controversial breakthroughs...
but is capable of extremely high quality world-class followup and large scientific investments given a clear objective target and government marching orders
V3 and R1 are impressive but didn't advance the absolute capabilities frontier. Maybe the capabilities/cost frontier, though we don't actually know how compute efficient OAI, Anthropic, GDM are.
I think this part of @gwern's comment doesn't hold up as well:
2. there is no interest or investment in an AI arms race, in part because of a "quiet confidence" (ie. apathy/lying-flat) that if anything important happens, fast-follower China can just catch up a few years later and win the real race. They just aren't doing it. There is no Chinese Manhattan Project. There is no race. They aren't dumping the money into it, and other things, like chips and Taiwan and demographics, are the big concerns which have the focus from the top of the government, and no one is interested in sticking their necks out for wacky things like 'spending a billion dollars on a single training run' without explicit enthusiastic endorsement from the very top.
I still don't think DS is evidence that "China" is racing toward AGI. The US isn't racing toward AGI either. Some American companies are, with varying levels of support from the government. But there's a huge gap between that and Manhattan Project levels of direct govt investment, support, and control.
However, overall, I do think that DS has gotten the CCP more interested in AGI and changed the landscape a lot.
I think that is a problem for the industry, but probably not an insurmountable barrier the way some commentators make it out to be.
- o-series of models may be able to produce new high quality training data
- sufficiently good reasoning approaches + existing base models + scaffolding may be sufficient to get you to automating ML research
One other thought is that there's probably an upper limit on how good an LLM can get even with unlimited high quality data and I'd guess that models would asymptotically approach it for a while. Based on the reporting around GPT-5 and other next-gen models, I'd guess that the issue is lack of data rather than approaching some fundamental limit.
Yup, discussed here: https://x.com/GarrisonLovely/status/1926095320997368319?t=vfuPigtomkOn5qc9Z8jCmQ&s=19
I think this is a more significant walkback than what I originally reported. Great find!