M

i like this because i can deeply apply research in alignment with edifying incentives even as an entrepreneur

@arram posterior value multiplier

To be fair I did an absurd amount of work to get access to my emotions so YMMV, but I'm pretty sure this will be incredibly useful for almost anyone.

Claude is better than the average therapist and 50x cheaper. It's like a sniper for avoided emotions. Unbelievably piercing insights that repeatedly have me in tears. Just surgical precision.

i should become a wagie

1. "open source" should not always mean "100% of our code is public 100% of the time" How much energy have we put into preventing .env leaks in our source control? How many miserable ways have we re-invented env var sharing? How many projects would be open source if they could hide in-flight PRs? How many security fixes are sitting unpublished because they will be exploited as soon as they appear in the tracker? How much better would life be if I could have a monorepo with some sub-packages that are "private" without splitting into multiple repos?

@BetaTomorrow despite this, I see your deep intentionality and poetics, and I really appreciate and resonate with it

@init_malachi English is not my native language, plus dyslexia, but point taken. Thanks for the feedback.

This paper is very strong because it focuses on a critical area: how sequence models store and retrieve relational structure in weights and embeddings. Its key contribution is to show that memory is not merely associative lookup; learned embeddings can organize into a geometric structure where unseen multi-hop relations become accessible. In Deep Manifold language, this paper studies the mathematical cover of memory: the learned weight/embedding geometry that defines possible fixed-point basins, relational neighborhoods, and latent intrinsic pathways. The limitation is that this paper mainly examines the static geometry of learned memory, not the dynamical activation process during inference. From the Deep Manifold view, inference is a boundary-conditioned iterated integral: the prompt acts as the boundary condition, and each layer produces a physical cover through activations. At every layer, the mathematical cover constrains what pathways are possible, while the physical cover selects, bends, and transports the current trajectory. Thus geometric memory is not simply stored geometry; it becomes effective only through repeated mathematical-cover / physical-cover interaction across stacked manifold layers.

i’m a question designer question engineer

@afterearthisnow Idk I think this might be partially one of the greatest purposes of ai It forces us to come to terms with what actually matters and actually choose our values very explicitly

Landian philosophy has reigned for so long because he's actually just one of the best living poets. Now that the pope is an equally good poet suddenly the cold god feels less inevitable. The power of art is very evident in this battle We probably shouldn't have defunded so much art and told everyone it's a waste to study humanities

the trick that makes @usekernel fast: we pay the full 5-10s of starting chromium once, snapshot ram to nvme, and resume in ~30ms when a request lands. from the outside it looks like a giant warm pool. under the hood it's mostly cold disk. wrote up how we got here.

I have been very impressed by @SemiAnalysis_ . I think of myself as a wide ranging systems engineer, looking for value at every level from the chip specs to the user interface, but SA exposes me to additional levels of "the system", both above (datacenters) and below (semiconductor fabrication). It probably puts me in "just knows enough to be dangerous" territory. Neat things I learned today: Some of the 800VDC datacenter design choices leverage parts commoditized by electric vehicles. There is now a SiC MOSFET that can operate on 10kV electricity, opening up the possibility of working directly with medium (ha!) voltage AC power transmission lines without stepping down.

To ensure our grading is fair and reliable, we built a trajectory analysis agent to replay agent rollouts and map out exactly why they fail. Running it on existing benchmarks surfaced significant grading noise, with verifiers rejecting valid code or letting models read solutions straight from git history.

@alokbishoyi97 i like it and i think it’s worth pursuing in kind in other areas

@init_malachi expand a bit ? what do you think fails - or issues with the DAG approach mentioned?

keep finding representation structures that elegantly hold processes and are expensive or non expressive within an LLM internal state

AI is NOT automated Logos, at least not yet. AI is automated Thumos. AI mechanistically relies through and through on mimetic imitation rather than a rational state machine solver. Part of the strangeness of its "learning" mechanism is it can approximate a logocentric style state machine after enough imitative repetitions. This understandably confuses many people. Before LLMs every person's experience with computer software was an encounter with hardcoded logos. Algorithms and state machines function at bottom by the programmer naming states and then gradually converting between states in the pattern until he can name the Tao/nature of his problem. AI neural nets solve problems by directly imitating the nature of the answer, letting its universal function infer the intermediate logos states.

There's a rule of thumb in Factorio: "avoid item/fluid buffers" which you only tend to learn after quite a while. (Buffers don't really save you, but do prolong the time before you find a broken process) I think about this a lot in the software industry.

I made a /seppuku skill for my Claudes for when they make an unforgivable mistake, and now they use it spontaneously without me asking.

@BetaTomorrow shadow is not as legible as light

@init_malachi no... make it balance

There is no pure “noise” per se. What the paper calls noise may simply be data the evaluator does not like, rare data, conflicting data, minority data, or data that does not fit the current measurement frame. From the Deep Manifold view, this is better understood as perturbation rather than noise. A neural network learns in a stochastic world, and stochasticity is not merely error; it is part of the inequality structure through which the model forms stochastic fixed points. Deep Manifold frames neural stochasticity as sum-based group statistics and boundary-shaped variability, not as a single bad disturbance. A small amount of such “noise” can be powerful because it perturbs the fixed-point trajectory without destroying the manifold. Often, a low ratio, say below roughly 5%, acts like useful boundary diversity: it prevents premature collapse, keeps nearby fixed-point classes reachable, and helps the model discover better convergence directions. The danger is not perturbation itself, but excessive or badly structured perturbation. Small perturbation stabilizes exploration; too much perturbation overwhelms the boundary condition and produces fixed-point drift. Paper: LLMs as Noisy Channels: A Shannon Perspective on Model Capacity and Scaling Laws Authors: Xu Ouyang, Deyi Liu, Yuhang Cai, Jing Liu, Yuan Yang, Chen Zheng, Thomas Hartvigsen, Yiyuan Ma. Affiliations: ByteDance Seed; University of Virginia; University of California, Berkeley arxiv.org/abs/2605.23901

@growing_daniel zing

You can tell AI is a net good for society because mark zuckerberg is bad at making it

@BetaTomorrow then filter

@init_malachi imbalance is bad

the one thing @mitsuhiko taught me: merged client & server logs. very useful.

this girl on tiktok said that “discernment is an olympic sport” and I’ve decided this is pretty much the driving sentiment of my life. it is Not all the same. all relationships are not equal. you Are responsible for your choices. your passivity is just carelessness