DeepSeek-R1 just released. Fully open source & transparent with MIT license. Developed with reinforcement learning directly on the base model. 20-30x cheaper API at comparable performance to OAI’s o1. (USD) 0.14 / million input tokens (cache hit). 0.55 / million input tokens (cache miss). 2.19 / million output tokens. To simplify, R1 is like R1-Zero but with multi-stage training: Its pipeline: • Fine-tune base with CoT data points • RL stage similar to R1-Zero • SFT using ~600k data points from rejection sampling and supervised datasets (e.g., writing, self-cognition). • RL stage to optimise objectives: helpfulness, harm reduction, etc. Emergent properties like longer responses, reflection, and alternative exploration emerge as natural products during training without explicit programming. RL rewards focus: • Accuracy (e.g., unit test-based scoring for code). • Format (e.g., tags for reasoning separation and language consistency). No outcome/process RMs used (simpler).

Whilst anthropomorphising AI using language like lies and so forth may be useful, e.g. for testing the "flavour" of the model, making the leap to becoming possessed with the idea that these things genuinely have some form of experiential intent and other things of that nature seems delusional, and baking that belief into training and models hyperstitionally could be a really dumb way to mess things up.

This is slop. It's also antithetical to some of the core tenets of Western civilisation. It is littered with contradictions to core ideas of Western philosophy and some of its greatest minds and the traditions held for thousands of years, from both the Hellenic (Plato, Aristotle, Plotinus) and Christian (St Irenaeus, St Athanasius, St Basil the Great, St Gregory of Nyssa, St John of Damascus) worldviews. Take for example, the video's notion of "built perfection". This is a decent example of such a contradiction and antithesis of classical Western philosophy and everything built upon it. Beyond that, it is also an internal contradiction in terms, and makes about as much sense as other contradictions in adjecto, such as a married bachelor or a square circle, where the qualifier negates the meaning of the noun and generates nonsense and falsity evident from the meaning of the terms themselves. It simply is wordslop. The definition and idea of perfection date to Plato and beyond (to Parmenides, for whom being is “in need of nothing” and “complete on every side” (Parmenides, fragment B8.33 and B8.42–43 Diels–Kranz)) and this idea is similarly present in the post-hellenic Christian revelatory worldview. Perfection as understood by Plato as in the sense of the forms, as completeness: the Form (of Beauty) is “beauty absolute, separate, simple, and everlasting, which without diminution and without increase, or any change, is imparted to the ever-growing and perishing beauties of all other things” (Symposium 211a–b); the good is “[t]he most perfect, Socrates, of all things” and “sufficient” “in a degree surpassing all other things” (Philebus 20d); and of the perfect, “If he change at all he can only change for the worse, for we cannot suppose him to be deficient either in virtue or beauty” (Republic II 381c). Further, this is clarified and explained by Leibniz as “perfection being understood as the magnitude of positive reality in the strict sense, when the limitations or the bounds of those things which have them are removed. There where there are no limits, that is to say, in God, perfection is absolutely infinite”, and “created things derive their perfections through the influence of God, but their imperfections come from their own natures, which cannot exist without limits” (Monadology §§41–42), and many arguments surrounding perfection, such as: “the necessary entity cannot lack any perfections as a lack of perfections means the existence of potentiality, as those lacks could be fulfilled. To be purely actual, all perfections and potentials must be realised and fulfilled”, and “composition of parts means potentiality and contingency as the entity is contingent on the set of parts’ arrangement (possibly) and the existence of all parts” (Muraviev, 2024; Muraviev, 2025a; Muraviev, 2026); “[w]e can derive the existence of being which is: Necessary, partless, singular, perfect.” (Muraviev, 2024); and anything built, being composed of parts and having passed from potentiality into being, cannot be (truly) perfect; “[a]nything created is, eo ipso, not-God and imperfect (in, at least, the God-sense)” (Muraviev, 2025b, cited in Muraviev, 2026). As such we see that to have "built perfection" makes no sense as a proposition. Furthermore, if we even were to suspend this contradiction, what would it imply about everything we are doing now, if we had already built perfection? Going beyond the contradiction and suspending this sort of AI-slop flattening of meaning and underlying content, if we were to take this as something humanly expressed, to avoid anthropomorphising code too much, if we were to assume the validity of the Christian tradition, this idea of having built perfection is, in itself, anti-Christian hubris, and fundamentally a babylonian notion, beyond even the alternative view coming from hellenists and vitalists such as BAP @bronzeagemantis, and their interpretation of Western civilisation. Fundamentally, the hellenic tradition and the tradition revealed after this via the development of Christian thought after the incarnation of Christ, and this pseudosynthesis (which is revelation instead of a synthesis in a constructed sense) shows a unified Western opposition to the validity of this idea. I think even those associating themselves with NRx, and those who subscribe to ideas of some form of ASI technological acceleration and singularity in a worshipful sense, would not even verily hold this proposition to be true. I think Nick Land @xenocosmography knows enough philosophy to also recognise that the hubris of attempting to "build" perfection (a different proposition to the even more preposterous claim of already having built it) simply makes no sense. I think this is fundamentally part of a delusion and distraction and misunderstanding of Western philosophy and both the hellenic and Christian position and going beyond this to more modern innovations and worldview; I do not think that even the furthest techno-optimists can sensibly defend the nonsensical notion of a temporally constructed "perfection", or defeat of entropy. And this sort of flattening, of compression of an impression of a complex idea, into something so oversimplified that the more complex underlying idea seems to be lost and in some sense overwritten and over-vectorised (planning to publish on this topic soon), seems to be a major part of AI risk which is possibly quite overlooked, especially when (to reference the meta-point of reinforcement, positive feedback RLHF) we give positive feedback to distortions and slop. If uncareful, there may seem to be a category of potential risk where AI and its surrounding systems may be sufficiently competent but insufficiently comprehending, and as such be destructive, if the instinct of discernment and preservation isn't drilled in. As such, I would caution against positive feedback for such slop. Of course, this may seem to be a disproportionate reaction to what is essentially a small, silly video produced by code running on silicon and metal (not to underestimate the immense scaling potential). However, I think the overall trend of oversimplification and ensloppification, and a sufficiently competent takeoff which magnifies these sorts of trends and follows those vectors excessively, could produce quite a regrettable state of affairs. Full post with references: victorm.co.uk/on-slop-and-we…

Used one of my usage resets earlier today in Codex. Still have 89% usage left; I understand that Chat & Codex have separate usage pools (why doesn’t Chat have usage tracking like Codex e.g. Pro usage left). Lost access to 6 Pro and was downgraded to lower quality models until tomorrow. I somewhat understand the situation with demand for Astra and the pause on new subscriptions to $200 Pro plans. I’ve been an OpenAI user for years before ChatGPT launched and love using your products. Perhaps consider additional usage reset purchase availability for Pro & Chat & clearer usage metrics (e.g. like in Codex)? Also please clarify whether failed Pro thinking attempts count against usage. @sama @nickaturley @thsottiaux @pvncher @vijayeraji @gdb @OpenAI @ChatGPTapp

The problem of substituting real analysis of OOD problems with simpler fakework strawmen is virtually intrinsic to current large larping models. Firstly, there’s a noumenological problem more broadly. However, that’s tangential to feasible short term progress, namely via further integration with evolving hardcoded harnesses. LLMs very frequently when corrected and given an instruction or rule (even a very simple one) fail to actually apply it even when they can clearly recite the rule. False completions and so forth, rule-forgetting, etc… are often harness failures. A personal example being a 5.6 Sol Ultra session (coordinator) instructing another 5.6 Sol Ultra (worker 2) to implement fixes exported from discussion with 5.6 Sol Pro (plan-level), with ultracode Fable 5 & Opus 5 auditing the process; however, what will often happen is that worker 2 gets the instruction yet simply doesn’t implement. This often gets missed without another manually prompted system check on that specific instruction. Very often even 5.6 Pro will simply “forget” to do things as basic as keeping the changelog, or not modifying forbidden/black text! I imagine that implementing the feature of evolving sub-harnesses would likely improve performance and compliance for LLMs/LRMs.

Bifurcation to permanent underclass seemingly postponed. :)