Sheikh Abdur Raheem Ali
3 followers · 11 following · 414 views
on the atlas · 51
- Some Memories of Early Anthropic — LessWrong1 savers
- dari.zone the island2 savers
- Tributary1 savers
- Middle Powers and Transformative AI2 savers
- 2.1.5 Authorization, Licensing, and Control @ Tracks1 savers
- Battle_of_the_sexes_(game_theory)2 savers
- Emile Durkheim, The Origin of the Notion of Totemic Principle, or Mana1 savers
- Dario Amodei — We Must Pace the Frontier37 savers
- Abdul Rahim Khan-i-Khanan1 savers
- Option Greeks explained: Delta, gamma, theta, vega, and rho | Wealthsimple1 savers
- Quote request: "if even the Sun requires proof" — LessWrong1 savers
- Understanding Data Center Power Delivery - Amodo Design Notes1 savers
- Verification Plan2 savers
- A Sketch of Good Communication — LessWrong1 savers
- The limits of AI safety via debate - LessWrong2 savers
- Lonely Dissent - LessWrong3 savers
- Singlethink — LessWrong1 savers
- Samsara | Slate Star Codex2 savers
- Falkenblog: One-Month Trading Strategies1 savers
- Notes on the Industry Job Search18 savers
- The Machines Lack Honour — LessWrong3 savers
- Fabien's Shortform — LessWrong1 savers
- The Financial Ledger Theory of Apologies — LessWrong1 savers
- Federal Register :: Framework for Artificial Intelligence Diffusion1 savers
- ryan_greenblatt's Shortform — LessWrong1 savers
- SSRIs: Much More Than You Wanted To Know | Slate Star Codex1 savers
- Auto-review of agent actions without synchronous human oversight3 savers
- GPU Driver Developer’s Guide — The Linux Kernel documentation1 savers
- Unreliable Guide To Locking — The Linux Kernel documentation1 savers
- VFIO - “Virtual Function I/O” — The Linux Kernel documentation1 savers
- Efficient LLM Finetuning with Unsloth | Modal Docs1 savers
- Dynamic batching | Modal Docs1 savers
- Bacteriophage1 savers
- I Wish People Were More Public8 savers
- Why Control Creates Conflict, and When to Open Instead — LessWrong1 savers
- Social Control Disorders - by Paola - Phenoatypical1 savers
- How Compiler Explorer Works in 2025 — Matt Godbolt’s blog2 savers
- Postcards | Club for the Future1 savers
- The machines are fine. I'm worried about us.23 savers
- Self-experiment risk-taking interview, by Elizabeth van Nostrand, Gwern · Gwern.net1 savers
- Intersection Observer API - Web APIs | MDN3 savers
- [AN #70]: Agents that help humans who are still learning about their own preferences — LessWrong1 savers
- Dario Amodei – The Adolescence of Technology — LessWrong1 savers
- Unsong1 savers
- On neural scaling and the quanta hypothesis12 savers
- What is the External Researcher Access Program? | Claude Help Center1 savers
- Pandoc - FAQs1 savers
- Dario Amodei — The Adolescence of Technology44 savers
- View article22 savers
- Reflections on 2025 - Samuel Albanie7 savers
- Mono no aware - Lightspeed Magazine3 savers
highlights · 44
We think the relative engineering and logistical costs involved in building new datacenter capacity on the ocean in international waters will be low enough due to robot and AI labor to make this worth it around 2035
Verification PlanThis is where you haven't changed your model, but decide to agree with the other person anyway.
A Sketch of Good Communication — LessWrongExcellent, I'm glad we've converged!
The limits of AI safety via debate - LessWrongthe difference between joining the rebellion and leaving the pack.
Lonely Dissent - LessWrongI had refused to play a negative-sum game.
Singlethink — LessWrongMaybe if they’ve created a super-efficient science of enlightenment, I would have to create a super-efficient science of samsara.
Samsara | Slate Star CodexHigh one-month idiosyncratic volatility predicts low returns.
Falkenblog: One-Month Trading StrategiesI failed my first behavioral interview because I went into it thinking I’m obviously well-“behaved,”
Notes on the Industry Job SearchIf you are going to create a system which takes morally significant actions, irrespective of whether it is a moral patient, then the main responsibility you incur — to it and to yourself and to the rest of the world — is to be good
The Machines Lack Honour — LessWrongI expect safety-related decisions during the intelligence explosion to look more like war-time decisions than risk assessments for nuclear power plants
Fabien's Shortform — LessWrongTaking responsibility for the costs you impose on others, and being a responsible leader of risky ventures, are natural and good, but will sometimes lead you to be responsible for bad outcomes you couldn't prevent and cannot rectify.
The Financial Ledger Theory of Apologies — LessWrongif data-generation model A is used to train data-generation model B and data-generation model C, and models B and C are used to train model D, then the operations to train A are only added to the number of operations for model D once.
Federal Register :: Framework for Artificial Intelligence Diffusion. If more than ten percent of the `operations' involve training on data that was not “published” as defined in § 734.7(a) and was generated by a single data-generation model, then `operations' the data-generation model used to generate the data should be counted, and if the data-generation model's `parameters' were not “published,” then the training `operations' used to train the data-generating model should be counted as well.
Federal Register :: Framework for Artificial Intelligence DiffusionNeuralese decoding prep: Make natural language autoencoders much better, build methods for extracting internal CoT, build better evaluations of how well natural language autoencoders work.
ryan_greenblatt's Shortform — LessWrongpeople’s feelings are the last thing to improve during recovery from depression.
SSRIs: Much More Than You Wanted To Know | Slate Star CodexTo reduce the likelihood of this occurring, we automatically stop the trajectory after repeated denials.
Auto-review of agent actions without synchronous human oversightyou probably need to be able to sleep, too
Unreliable Guide To Locking — The Linux Kernel documentationDMA is by far the most critical aspect for maintaining a secure environment as allowing a device read-write access to system memory imposes the greatest risk to the overall system integrity.
VFIO - “Virtual Function I/O” — The Linux Kernel documentationa common estimate for naive finetuning puts the VRAM requirements at roughly 4.2x the original model size: 1x for model weights + 1x for gradients + 2x for optimizer state + 20% for activations
Efficient LLM Finetuning with Unsloth | Modal DocsBatching increases throughput at a potential cost to latency.
Dynamic batching | Modal DocsΦ3T makes a short viral protein that signals other bacteriophages to lie dormant instead of killing the host bacterium.[86] Arbitrium is the name given to this protein by the researchers who discovered it
BacteriophageMy default mode is solipsism. I read in private, build in private, learn in private. And the problem with that is self-doubt and arbitrariness. I’m halfway through a textbook and think: why? Why am I learning geology? Why this topic, and not another? There is never an a priori reason.
I Wish People Were More Publiccontrol often intensifies the states it tries to suppress.
Why Control Creates Conflict, and When to Open Instead — LessWrongI suspect that quite a lot of mental health issues develop to control, rather than to adapt.
Social Control Disorders - by Paola - PhenoatypicalLLM also reminded me that I usually put a disclaimer at the end of my articles saying that I used AI assistance, which I had forgotten to do. Thanks, Claude.
How Compiler Explorer Works in 2025 — Matt Godbolt’s blogSend your postcard to Club for the Future, we'll launch it to space and back on a New Shepard rocket, stamp it "Flown to Space," and return it to you.
Postcards | Club for the FutureClaude had been adjusting parameters to make plots match instead of finding actual errors. It faked results. It invented coefficients. It produced verification documents that verified nothing. It asserted results without derivation. It simplified formulas based on patterns from other problems instead of working through the specifics of the problem at hand.
The machines are fine. I'm worried about us.Andrew Gelman has a rule of thumb that to detect an interaction needs something like 16× more data than to detect just the main effects.
Self-experiment risk-taking interview, by Elizabeth van Nostrand, Gwern · Gwern.netTo get a feeling for how thresholds work, try scrolling the box below around.
Intersection Observer API - Web APIs | MDNThe state controllability condition implies that it is possible – by admissible inputs – to steer the states from any initial value to any final value within some finite time window. A continuous time-invariant linear state-space model is controllable if and only if rank [ B A B A 2 B ⋯ A n − 1 B ] = n , {\displaystyle \operatorname {rank} {\begin{bmatrix}\mathbf {B} &\mathbf {A} \mathbf {B} &\mathbf {A} ^{2}\mathbf {B} &\cdots &\mathbf {A} ^{n-1}\mathbf {B} \end{bmatrix}}=n,} where rank is the number of linearly independent rows in a matrix, and where n is the number of state variables.
State-space representationHowever, in cooperative settings, things are not so nice: a failure to anticipate your partner's plan can lead to arbitrarily bad outcomes.
[AN #70]: Agents that help humans who are still learning about their own preferences — LessWrongif the free dimension is in all inputs it’s a batch dimension, and if it’s missing from some inputs we will broadcast those tensors.
Computing sharding with einsum : ezyang's blogThen there will be a time for courage, for enough people to buck the prevailing trends and stand on principle, even in the face of threats to their economic interests and personal safety.
Dario Amodei – The Adolescence of Technology — LessWrongThen there will be a time for courage, for enough people to buck the prevailing trends and stand on principle, even in the face of threats to their economic interests and personal safety.
Dario Amodei — The Adolescence of TechnologyThen there will be a time for courage, for enough people to buck the prevailing trends and stand on principle, even in the face of threats to their economic interests and personal safety.
Dario Amodei — The Adolescence of TechnologyYet it is this awareness of the closeness of death, of the beauty inherent in each moment, that allows us to endure. Mono no aware, my son, is an empathy with the universe. It is the soul of our nation. It has allowed us to endure Hiroshima, to endure the occupation, to endure deprivation and the prospect of annihilation without despair.
Mono no aware - Lightspeed Magazinesignificantly reduce reward confusion by leveraging transitivity of preferences while building a global preference chain with active learning.
View articleThe kabbalists are only trying to understand the world. The point is to change it.
UnsongOn the right, we see a separate cluster of newlines being predicted by the model. What these samples have in common is that they are newlines in line-length-limited text. This corresponds to the skill of counting the length of lines, and then predicting a newline to maintain the length of the previous lin
On neural scaling and the quanta hypothesisPlease note that given the substantial number of applications we receive (sometimes thousands in a single week), we regret that we cannot provide individual responses to unapproved submissions.
What is the External Researcher Access Program? | Claude Help CenterI used pandoc to convert a document to ICML
Pandoc - FAQsTo evaluate code you need an engineer capable of reviewing a diff without succumbing to the urge to rewrite it in Rust.
Reflections on 2025 - Samuel AlbanieThe pstree subcommand fetches the same information, but instead renders a visual async call tree, showing coroutine relationships in a hierarchical format. This command is particularly useful for debugging long-running or stuck asynchronous programs.
What’s new in Python 3.14 — Python 3.14.2 documentationA very fast mouse event will have a high probability of being missed on a low sampling rate, but will always be collected as a marker.
Profiler Fundamentals