Hot take: Please, do not make yourselves look like extremists.

Or: be aware that bad press does exist, and it might be harmful to the overall cause to provide quotable resources for adversaries; and thus refrain from calling a possible invasion of Taiwan an "act of God", at least not with your public account with your name on it, as one of the top five or so most prominent alignment figures right now, to be recorded in the Internet Archive for all posterity.

(Note that it's not just "don't be extremists", but also "don't look like one", which, unfortunately, is more of a PR than a morality problem. I have good faith that AI safety people mean well, and Daniel, with his positive, admirably considerate, and moderate vision for AI 2040, seems very unlikely to mean ill. This might heckle some very powerful people, but I think it's important to say clearly.)

As Eliezer said, only law can prevent extinction, and I struggle to believe that it'd provide the community with more respect or trust from the establishment, for it to be one of the few entities that actively want a war that might kill the most American troops since Vietnam, or "Hugging Face 2: This Time in a Hospital". (Sources: https://taskandpurpose.com/news/us-china-war-casualties/, https://ipdefenseforum.com/2026/01/pla-would-suffer-100000-casualties-in-failed-taiwan-invasion-study-says/.)

Call me a pessimist or inexperienced, but it seems very plausible to me that, even if we do indeed convince the establishments worldwide that AI safety is important and ASI safety is paramount, the proceeding effort to convince that fundamental alignment is more important than prosaic alignment would depend heavily on the perceived trustworthiness and morality of each involved faction.

It is plausible - though a stretch - for the world to forgo prosperity in favor of security; and certainly the arguments about prosaic alignment's insecurity have enough merit that cautious, self-preserving legislators might lean in favor of them, all else excluded. Even if that might slow down AI progress and forgo the numerous economic, military, and subsequent political benefits from it.

But that's "all else excluded"; if the perception of fundamental alignment theorists is that they are extremists, then that perception might undermine the security benefits - arguably the only thing we can offer the establishments over the existing approaches by Anthropic or OpenAI - perceived by legislators when trying to choose which experts to trust (and tip the balance back in favor of a less desirable faction).

Abdul El-Sayed, senator-elect of Michigan, has called for a full-on AI pause and treaty with China: https://x.com/ONagel33303/status/2117452882773684550. It's interesting that he also echoed policies pushed forward by AI 2040 and Pacing The Frontier, such as an international treaty with China to prevent an arrival at RSI (which was mentioned by name and metaphor in the interview) and the "preservation of humanity" (word for word), and adopted positions that loosely seem to mean that China's progress is driven by American progress (albeit the mechanism, distillation, was not outright mentioned).

Though an increasingly popular position among the left wing of the Democrats, one could perhaps take heart at such statements by figures of increasing prominence in support of the cause. The tweet also received 4.9 million views on Twitter, possibly showing somewhat growing interest in AI safety.

Speaking as somebody from the Global South, IMO, Thinking Machines (which makes Inkling) is one of the most geopolitically important labs in the United States right now, just behind the big 3: OpenAI, Anthropic, and GDM. The reason is that, for over a year since gpt-oss-120b's release, US AI progress has not been correlated with global benefits in independent AI research, namely, that they do not advance open-source models. For this reason, behaviors that were obviously unethical and undesirable (distillation, Chinese AI advancement) were justified by many around the world as acceptable or even noble in some cases, as they're the only supplier of open-source models that are so vital to any sort of AI research.

If Inkling continues its trajectory of rapid growth (and if the open-source ecosystem in America and Europe grows enough to rival China's), I'd expect global sympathy for Chinese methods to drop reasonably quickly, especially given significant security concerns around Chinese models. Whether or not this is a good thing remains up for debate; on one hand, a larger lead over China (not just in pure capabilities but in demonstrable commercial numbers) could help with pressuring them, and at least many countries would switch to using American open-source models as a template to develop their own sovereign systems, as well as many researchers would switch. On the other hand, it might provoke more aggressive Chinese measures as they feel cornered, and plausibly weaken an international deal. Nonetheless, as the preeminent open-source lab in America, Thinking Machines should be considered a rather important lab.

(This is specifically in response to the release of Inkling 2 Small's weights yesterday, which are estimated in some sources (AA) as nearly GLM-5.3 levels and better than Mistral Large 4. I am still waiting for its ECI numbers.)

Current high-level harnesses can uplift performance by ~7-12 ECI points; therefore, IMO safety thresholds (especially post-TED-AI or SAR) should be lowered by at least 10 ECI points.

AI co-mathematician, a "harness of harnesses" developed by GDM back in May 2026, achieved a performance in FrontierMath v4 in the ballpark of several models at 160-164 Math ECI. (These are Gemini 4 Argon (165), GPT-5.5 Pro (164), Claude Opus 5 (164), GPT-5.5 (160), and Muse Wonder (162); it performs below the first 2 and above the last 3, with Comathematician's results on the benchmark higher than the mean of the 5 models.)

The mean Math ECI of these models is 163, which is a full 10 ECI points above the Math ECI of Gemini 3.1 Pro, which powered Comathematician (153). It is likely non-trivial to scale sophisticated harnesses upwards across various levels of capability; otherwise, we'd probably have observed enormous mathematical results powered by enhanced Gemini 4 Carbon.

Nonetheless, I think this is a concrete display that massive capability leaps can happen from harnesses only, and that such harnesses were designed by human-level intelligences, an effort that could likely be replicated (plausibly even trivially) by SAR/TED-AI level models to spike a misaligned model's capabilities, without spiking model capabilities. This is especially relevant in cases like Plan C, where an underestimation of pre-ASI models' abilities (or their ability to strategically spike beyond a dangerous threshold, even if inefficiently compared to weak or strong RSI) could lead to very dangerous outcomes.