The intake queue
The reading list
Everything I saved, and what I decided to do with it. Each link gets exactly one ruling — read it, propose it for the Atlas, write something, build something, or bin it. The ones I binned are on this page too: a list that only showed the keepers would make the judgement look better than it is.
LAST TRIAGE
The open queue
17 items · grouped by what I said I'd do with themExplainer 1
- Queued22580: From GPT2 to Kimi3, Explained ↗
22,580 GPT-2 (2019) models fit inside Kimi K3 (2026) — a 22,580x scale-up in seven years. Asks whether the gains are 'just... scale?'
@waterloo_internExplainerShortOnly read about itmodelsIn the content queueSitting 2+ weeks
My take on their work 1
- QueuedLilian Weng (OpenAI) on Harness Engineering ↗
Deep breakdown of harness design patterns, workflow automation, persistent memory via file systems, and how modern agents like Claude Code actually work. Recommended as the best blog post of the month.
@micahomndProposed for the AtlasMy take on their workOnly read about itharnessesIn the content queueSitting 2+ weeksThe thing itself ↗
Proposed for the Atlas 3
- QueuedModular's LLM Inference Handbook ↗
Vendor-authored reference covering tokenization, attention, prefill/decode, speculative decoding, continuous batching, prefill-decode disaggregation, and performance benchmarks. The same post quote-tweets Google DeepMind's 'How to Scale Your Model' (jax-ml.github.io/scaling-book).
@vivekgalatageProposed for the AtlasOnly read about itinference-economicsSitting 2+ weeksThe thing itself ↗ - QueuedContinual learning is an ICL problem, not a weights problem ↗
'Reading this paper changed my priors on continual learning.' Quotes Charlie O'Neill (@oneill_c): all the training that creates a model is engineered around crafting the best possible in-context-learning mechanism, and further training DEGRADES that mechanism — so continual learning should focus on loading context rather than updating weights. O'Neill's article: 'Trying to actually define continual learning' — 'Every AI agent today is an amnesiac intern.'
@waterloo_internProposed for the AtlasOnly read about itharnessesSitting 2+ weeksThe thing itself ↗ - QueuedQuantization: error(a) + error(b) < error(a) alone ↗
Argues every AI company including Nvidia was needlessly sacrificing model intelligence and speed. Counterintuitive claim: quantizing MORE of the model lowers total error. Quotes @the_joshua_hill's paper — a SOTA quantization method using Fourier Analysis on Groups, achieving 20% higher throughput on GLM 5.2 vs existing configs.
@waterloo_internProposed for the AtlasOnly read about itinference-economicsSitting 2+ weeks
A question to chase 1
- QueuedThe routing problem birthed every flavour of sparse attention ↗
Claims the routing problem in ML inference is the single most unintuitive thing to solve for, and that it alone is responsible for the birth of every flavour of sparse attention — SWA, NSA, DSA.
@waterloo_internRead it, then decideA question to chaseOnly read about itmodelsSitting 2+ weeks
Read it, then decide 11
- QueuedTowards infinite context windows: neural KV cache compaction ↗
Neural KV cache compaction toward unbounded context. Surfaced 2026-08-06 while resolving the O'Neill bookmark — same research team.
Baseten ResearchRead it, then decideOnly read about itharnessesSitting 2+ weeks - QueuedRepeated KV cache for long-running agents ↗
KV cache reuse for agents that run a long time. Surfaced 2026-08-06 while resolving the O'Neill bookmark — same research team.
Baseten ResearchRead it, then decideOnly read about itinference-economicsSitting 2+ weeks - QueuedContinual learning and the post monolith AI era ↗
The cost of continual learning scales with generality. Surfaced 2026-08-06 while resolving the O'Neill bookmark — same research team.
Baseten ResearchRead it, then decideOnly read about itharnessesSitting 2+ weeks - QueuedAmnesiac generalist behemoths are not the future of language models ↗
You don't need a generic genius. You need a specialist learner. Surfaced 2026-08-06 while resolving the O'Neill bookmark — same research team.
Baseten ResearchRead it, then decideOnly read about itmodelsSitting 2+ weeks - QueuedRLHF & Post-Training course (Nathan Lambert) ↗
13 video lectures plus Q&A and guest talks, following the chapters of Lambert's RLHF book. L0 ML foundations; L1-4 core concepts and optimisation; L5-13 reasoning models, DPO, synthetic data, preference data, over-optimisation, regularisation, tool use, evaluation, character training. Slides, PDFs and markdown source published alongside the videos.
Nathan LambertRead it, then decideOnly read about itmodelsSitting 2+ weeks - QueuedMegakernels are dead ↗
Post-mortem on a Latent.Space podcast appearance — 'i'm sorry that i was right about every single thing. a) re megakernels are dead...'. Links the Latent.Space Inference Engineering Masterclass with Baseten/@philipkiely: 10x faster models, quantization, speculative decoding, Rubin, and self-optimizing AI.
@waterloo_internRead it, then decideOnly read about ithardwareSitting 2+ weeksThe thing itself ↗ - QueuedLiquid AI's small-model recipe ↗
Recommends Maxime Labonne's presentation. Quotes @h100envy: Liquid AI's head of post-training built an on-device model under 1 GB in 20 minutes — LFM2.5 base -> on-policy preference alignment -> agentic reinforcement learning -> curriculum training.
@paulabartabajo_Read it, then decideOnly read about itmodelsSitting 2+ weeks - QueuedDesigning and post-training small LMs (Cohere Labs Summer School) ↗
A 30-minute recorded presentation plus 40 slides on designing and post-training small language models.
@maximelabonneRead it, then decideOnly read about itmodelsSitting 2+ weeks - QueuedSuperposition: near-orthogonal directions beyond dim count ↗
'(friendly fire) every time i've disagreed with ml researchers i've been proven wrong, but i'm going to do it, just one more time' — weights hold near-orthogonal directions exponentially beyond dimension count, if within noise tolerance.
@waterloo_internRead it, then decideOnly read about itmodelsSitting 2+ weeks - QueuedRMSNorm is not actually cheaper than LayerNorm ↗
'Every research team needs to spend some time learning how their modeling code lowers down to gpu kernels. Example: rms norm vs layernorm. Industry assumes rms norm is cheaper. You don't need the x-hat, it looks simpler, so it must be faster, so it got adopted. But that's not true.'
@waterloo_internRead it, then decideOnly read about ithardwareSitting 2+ weeks - QueuedSome notes on writing the fastest video kernel in the world ↗
Worklog: retrofitted sparsity into a model at inference time, yielding what the author claims is the world's fastest video generation kernel — iteratively optimizing a frontier OSS model by 54x.
@waterloo_internRead it, then decideOnly read about ithardwareSitting 2+ weeks
What became work
3 promoted into the content queue- DroppedPower is the central bottleneck for AI, not chips ↗
A former Citadel quant who covered power & gas argues the bottleneck framing is wrong: 'There's constant talk about chips & memory, but power is the central bottleneck for AI. Very few people understand it.' Published a primer on power pricing and datacenters.
@neelsomaniRuled outOnly read about itdatacentersIn the content queueThe thing itself ↗ - QueuedLilian Weng (OpenAI) on Harness Engineering ↗
Deep breakdown of harness design patterns, workflow automation, persistent memory via file systems, and how modern agents like Claude Code actually work. Recommended as the best blog post of the month.
@micahomndProposed for the AtlasMy take on their workOnly read about itharnessesIn the content queueThe thing itself ↗ - Queued22580: From GPT2 to Kimi3, Explained ↗
22,580 GPT-2 (2019) models fit inside Kimi K3 (2026) — a 22,580x scale-up in seven years. Asks whether the gains are 'just... scale?'
@waterloo_internExplainerShortOnly read about itmodelsIn the content queue
Ruled out
8 binned, and kept on the recordA drop stays on the list. Partly so it is never offered to me a second time, and partly because what someone declined to read is as much a statement of judgement as what they kept.
- DroppedPower is the central bottleneck for AI, not chips ↗
A former Citadel quant who covered power & gas argues the bottleneck framing is wrong: 'There's constant talk about chips & memory, but power is the central bottleneck for AI. Very few people understand it.' Published a primer on power pricing and datacenters.
@neelsomaniRuled outOnly read about itdatacentersIn the content queueThe thing itself ↗ - DroppedSpeed Above All Else ↗
'Founders: print this out and put it on the wall.' Quotes Alfred Lin's article on founder velocity, referencing @tanay_tandon of Commure.
@alexisohanianRuled outOnly read about it - DroppedTurn any website into an API or MCP server (1/2) ↗
Technique for generating an API/MCP client against any website: capture the real network traffic in DevTools, then 'Copy all as HAR' and 'Copy all as fetch' to generate a client.
@anaisbettsRuled outOnly read about it - DroppedTurn any website into an API or MCP server (2/2) ↗
Continuation of the DevTools HAR technique.
@anaisbettsRuled outOnly read about it - DroppedSystems over goals ↗
Quote-card: 'If you do something every day, it's a system. If you're waiting to achieve it someday in the future, it's a goal.'
@thebeautyofsaasRuled outOnly read about it - DroppedDavid Sacks on Aschenbrenner's thesis ↗
Sacks summarising Aschenbrenner's three areas: raw compute, algorithmic efficiency, and unhobbling.
@firesidealphaRuled outOnly read about it - DroppedWeekend reading list ↗
Recommended weekend reading — Gavin Baker's pinned post and others.
@ParadisLabsRuled outOnly read about it - DroppedSATIRE — 'distilled 2.3M Claude Fable 5 traces into Qwen3-4B' ↗
Claims 100% self-consistency @ 512 samples, 0.00 bits output entropy, and zero hallucination variance — 'turns out the student is not bounded by the teacher.'
@waterloo_internRuled outOnly read about it
There is no tick-box here, for the same reason there is none on the curriculum: a second way to mark something done is a second version of the truth, and it would diverge within a week. The file docs/context/reading-list.json is the checklist, this page is a read of it, and the state on every row is that file's own. Freshness arrives with a deploy, not on a timer. This is the intake queue and it is a different thing from the 165-paper bibliography. The bibliography →