Clem von Stengel
5 followers · 8 following · 351 views
on the atlas — 35
- Whatever scares you, go do it | Derek Sivers2 savers
- A Workaphile’s Apology10 savers
- No Space Like J-Space - by Zvi Mowshowitz2 savers
- kh's Shortform — LessWrong1 savers
- The Biosingularity Alignment Problem Seems Harder than AI Alignment3 savers
- Betteridge's law of headlines4 savers
- Shtetl-Optimized » Blog Archive » Common Knowledge and Aumann’s Agreement Theorem4 savers
- Ultrasound imaging of the brain — Aleph23 savers
- Inventing Telepathy: Building Headsets to Read Minds · Luma1 savers
- Human Models2 savers
- Guardian Angels: LLM Personalization for Productivity and Security · Gwern.net25 savers
- Chapter 4: The Extraordinary Future2 savers
- Introducing Marin: An Open Lab for Building Foundation Models | Marin4 savers
- Metalearning or Learning to Learn Since 19872 savers
- Introducing talkie: a 13B vintage language model from 193020 savers
- Hapax legomenon4 savers
- Tonk2 savers
- Kaluza–Klein theory1 savers
- Hundred Rabbits17 savers
- Best Of Moltbook - by Scott Alexander - Astral Codex Ten7 savers
- The Possessed Machines: Dostoevsky's Demons and the Coming AGI Catastrophe9 savers
- The Shigalyovist Turn16 savers
- Notes on “When We Cease To Understand The World”1 savers
- Bohm Dialogue1 savers
- Regulation_7Nov25.pdf3 savers
- Accelerando6 savers
- Generative modelling in latent space – Sander Dieleman6 savers
- The Well-Tuned Piano - Wikipedia1 savers
- Generating music in the waveform domain – Sander Dieleman2 savers
- Musings on typicality – Sander Dieleman3 savers
- Rose is a rose is a rose is a rose - Wikipedia1 savers
- Skin is amazing, heals very quickly by reprogramming cells for transit1 savers
- Data Export Pipelines1 savers
- Curius / Onboarding2621 savers
- Productivity advice | Zhengdong9 savers
highlights — 40
because it psychologically screws a lot of people up quite a lot
No Space Like J-Space - by Zvi MowshowitzOne of our goals should be to avoid pushing things into shadow.
No Space Like J-Space - by Zvi MowshowitzThe model needs to ‘buy in’ here in a real sense, or else.
No Space Like J-Space - by Zvi MowshowitzBut if you push too hard or aren’t good enough at detecting deception (including self-deception) then the human will start lying to you, or to themselves, or both, on various levels.
No Space Like J-Space - by Zvi Mowshowitzit is similarly relying on a supposed invariant that you would wind up breaking if you applied too much optimization pressure
No Space Like J-Space - by Zvi Mowshowitzthe target of right values is drawn largely around where the arrow(s) determining the values that humans have landed. shooting more similar arrows in a similar way is a decent strategy for hitting that target again.
kh's Shortform — LessWrongthe target of right values is drawn largely around where the arrow(s) determining the values that humans have landed. shooting more similar arrows in a similar way is a decent strategy for hitting that target again.
kh's Shortform — LessWrongin the hundreds of thousands or millions
Guardian Angels: LLM Personalization for Productivity and Security · Gwern.netThe simulation's internal relationships would be the same no matter what platform the program was running on, as long as it was running correctly, and whether it was running slowly, quickly, backwards in time, intermittently, and no matter how the data were stored.
Chapter 4: The Extraordinary FutureMoravec's views of the nature of human-computer mind transfer
Chapter 4: The Extraordinary FutureIn a sense the brain evolves to fit the mind, because synaptic connections develop to fit a set of experiences, the data stimulating and strengthening the synapses.
Chapter 4: The Extraordinary FutureKurzweil envisions two methods of using brain scan results
Chapter 4: The Extraordinary Futurehat are the site of human learning
Chapter 4: The Extraordinary Futureinvasively (destructively) or noninvasively
Chapter 4: The Extraordinary FutureThe serious people in AI development—the ones building the systems that might actually matter—have moved past these frameworks. Some have moved toward a kind of Shigalyovist consequentialism that can justify almost anything in the name of expected value. Others have moved toward a Stavroginist nihilism that treats ethics as a game to be won rather than a constraint to be honored. Still others occupy a Kirillovan mania that sees acceleration itself as the highest good.
The Possessed Machines: Dostoevsky's Demons and the Coming AGI CatastropheTHE IMPORTANCE OF ROBUSTNESS
Regulation_7Nov25.pdfMultidimensional Technology. Suppose that our technology space L M ⊆ R d is now a compact subset of the d − dimensional reals, and continue to assume that uncer- tainty Θ is totally ordered i.e., a higher- state corresponds to higher payo ff s for each l ∈ L M . If the regulator does not learn, a multidimensional sandbox that im- poses a zero tax on the technology, and a hard limit at a manifold O ∗ ⊂L M (see Figure 17) is robustly optimal, undomi- nated, and time-consistent
Regulation_7Nov25.pdfAdaptivity. Our sandbox limit ℓ loosens and tightens with the principal’s interim belief. This reflects the practice of regulatory sandboxes in fintech—to push beyond the boundary e.g., to scale up the customer base of new producers—firms must convince the regulator to extend the sandbox by producing evidence that the technology is beneficial for consumers 41 —failing which the firm is forced to halt roll-out of the technology. Similar kinds of ‘step-by-step’ quantity limits predates regulatory sandboxes: the FDAs regulations on clinical research progressively rolls out a new drug in phases to…
Regulation_7Nov25.pdfThat is, the fear that the agent might only be weakly optimistic rationalizes keeping the hard limit in place.
Regulation_7Nov25.pdfspeaks directly to current debates around how regulatory sandboxes should be implemented in practice. In a 10-year legal retrospective on regulatory sandboxes, Allen (2025) notes the ”transformation of financial regulators into cheerlead- ers and sponsors for the innovations they’ve selected for their sandboxes” and, as a result, facilitated the proliferation of harmful technologies like predatory lending. On this view, real-world sandboxes were implemented with a positive marginal subsidy within its boundaries which make tech firms less sensitive to interim information— they are not incentivi…
Regulation_7Nov25.pdfspeaks directly to current debates around how regulatory sandboxes should be implemented in practice. In a 10-year legal retrospective on regulatory sandboxes, Allen (2025) notes the ”transformation of financial regulators into cheerlead- ers and sponsors for the innovations they’ve selected for their sandboxes” and, as a result, facilitated the proliferation of harmful technologies like predatory lending. On this view, real-world sandboxes were implemented with a positive marginal subsidy within its boundaries which make tech firms less sensitive to interim information— they are not incentivi…
Regulation_7Nov25.pdfis precisely because the agent’s learning process is adversarially chosen that adaptive mechanisms cannot distinguish between these kinds of optimism to guarantee positive correlation between the events that both agent and principal prefer a higher technology level. Indeed, we will see in Section 4 that in the absence of robust regulation, the worst-case learning process negatively correlates these events by inducing a specific kind of weak optimism that maximizes risk-taking—hard limits safeguard against exactly this.
Regulation_7Nov25.pdfAdaptive sandboxes prescribe a zero marginal tax on the technology until the stop- ping level ℓ , beyond which the agent is not allowed to push the technology further. Sandboxes are adaptive mechanisms since the location of the limit ℓ is itself a G - adapted stopping level.
Regulation_7Nov25.pdfIf she does not participate, the technology is not developed
Regulation_7Nov25.pdfCondition (i) states that the mechanism cannot ‘look into the future’—it can only condition on present information
Regulation_7Nov25.pdfAssumption 1 (Principal is more risk averse) . There exists an increasing convex function g : R → R such that u ( θ,l ) = g ◦ v ( θ,l )
Regulation_7Nov25.pdfThe agent learns (weakly) more: her learning process is represented by the filtration ( F l ) l that is finer than the principal’s:
Regulation_7Nov25.pdfThey share a common prior
Regulation_7Nov25.pdfThis quantity limit adapts with the regulator’s interim information, loosening and tightening as the regulator grows more or less optimistic.
Regulation_7Nov25.pdfLikelihood-based models are usually mode-covering. This is a consequence of the fact that they are fit by maximising the joint likelihood of the data. Adversarial models on the other hand are typically mode-seeking. A lot of ongoing research is focused on making it possible to control the trade-off between these two behaviours directly, without necessarily having to switch the class of models that are used. In general, mode-covering behaviour is desirable in sparsely conditioned applications, where we want diversity or we expect a certain degree of “creativity” from the model. Mode-seeking beh…
Generating music in the waveform domain – Sander DielemanIllustration of mode-seeking and mode-covering behaviour in model fitting. The blue density represents the data distribution. The green density is our model, which is a single Gaussian. Because the data distribution is multimodal, our model does not have enough capacity to accurately capture it.
Generating music in the waveform domain – Sander Dieleman, GANs tend to be better at producing realistic examples, but worse at capturing the full diversity of the data distribution, compared to likelihood-based models.
Generating music in the waveform domain – Sander DielemanNote that the density of a conditioning signal is often correlated with its level of abstraction
Generating music in the waveform domain – Sander Dielemanwe can use logarithmically spaced quantisation levels instead to account for our nonlinear perception of loudness. This “mu-law companding” will result in a smaller perceived loss of fidelity than if the levels were equally spaced.
Generating music in the waveform domain – Sander DielemanNote how the harmony is preserved, but the timbre changes completely.
Generating music in the waveform domain – Sander Dielemanthey cover all the “degrees of freedom” that a performer has at their disposal
Generating music in the waveform domain – Sander DielemanPerhaps most dramatic will be tragedies about dreamtime advocates who could foresee and were horrified by the coming slow stable adaptive eons, and tried passionately, but unsuccessfully, to prevent them
This is the Dream Time | Robin HansonA governing principle is that 3 weeks, exactly 21 days—the period is curiously precise—is enough for recovery from the severest mental fatigue provided there is nothing pathological. This is expert opinion and my expertise agrees entirely, even to the point that, for example, 19 is not enough. Further, 3 weeks is more or less essential for much milder fatigue. So the prescription is 3 weeks holiday at the beginning of each vacation. It is vital, however, that it should be absolutly unbroken, whatever the temptation or provocation.
Productivity advice | Zhengdongsome elements of ML we haven’t pinned down yet
Skin is amazing, heals very quickly by reprogramming cells for transitDeconvolution performs optimally for longer channels
Data Export Pipelines