flâneur

Ratan Kaliani

78 followers · 30 following · 5105 views

on the atlas — 533

highlights — 1595

  • We argue that in multi-turn RL, the frozen agents’ responses act as few-shot prompts that both influence future actions and serve as a distillation signal, helping the model quickly learn the game. Over time, the model learns exploitative techniques through heavy exposure to its competing models. This is a working explanation from our hackathon findings and we encourage research in the area!
    Natural Deception with RL - Rajan Agarwal
  • Conditional computation, where parts of the network are active on a per-example basis, has been proposed in theory as a way of dramatically increas- ing model capacity without a proportional increase in computation. In practice, however, there are significant algorithmic and performance challenges. In this work, we address these challenges and finally realize the promise of conditional computation, achieving greater than 1000x improvements in model capacity with only minor losses in computational efficiency on modern GPU clusters. We in- troduce a Sparsely-Gated Mixture-of-Experts layer (MoE),…
    [1701.06538] Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer
  • Some farmers are determined to ride the berry boom without being pulled into Driscoll’s ever more powerful orbit. Taylor Doyle, a third-generation fruit farmer in Western New York, just pulled out 95 acres of mature apple trees to make way for blueberry bushes.
    Why Are Berries Everywhere, in Every Season? Driscoll’s. - The New York Times
  • Chinese growers have succeeded because of access to advanced technologies like robots that spray pesticides and automated lighting that prevents mold, he said. “In the U.S., that would cost seven or eight times more,” he said.
    Why Are Berries Everywhere, in Every Season? Driscoll’s. - The New York Times
  • In 2025, China overtook the United States as the world’s largest blueberry producer. Driscoll’s, the first foreign berry company allowed to operate there, now has about 8,000 acres under cultivation.
    Why Are Berries Everywhere, in Every Season? Driscoll’s. - The New York Times
  • In 2023, they became Mexico’s most lucrative agricultural export, surpassing avocados, beer and tequila. On Moldovan plantations and in Andean highlands, growers of low-margin crops like sugar cane and corn have switched to berries, which command premium prices.
    Why Are Berries Everywhere, in Every Season? Driscoll’s. - The New York Times
  • Most of that growth has been driven by Driscoll’s, a $7 billion California company that began as a multifamily farm in 1904, patented its first strain of strawberries in 1958 and is still controlled by family members. In 1989, its board made what the company calls the Meadowood Declaration, a resolution that seemed preposterous at the time: to make all four berries available, in every season, in every part of the world.
    Why Are Berries Everywhere, in Every Season? Driscoll’s. - The New York Times
  • This “on-policy” background data sampled by Qwen3-8B acts as a forwards KL regularizer, reinforcing the model’s original behavior throughout mid-training. We find sampling from Qwen3-8B is better than Qwen3-32B for preserving chat capabilities throughout mid-training, highlighting the sensitivity of the data source; similar on-policy SFT results have been found in Chen et al.Retaining by Doing: The Role of On-Policy Data in Mitigating Forgetting (Chen et al, 2025) We hypothesize this approach can be even more effective than having access to the original pretraining data distribution, at the co…
    On-Policy Distillation - Thinking Machines Lab
  • Weakness mining: cluster failures into verifier-grounded failure patterns. The current harness is used to evaluate on tasks and execution traces are collected for analysis. Note that two runs can share the same verifier outcome in the error logs on the surface, such as timeout or missing artifact, while having different causal mechanisms. Therefore we need a failure record of rich information, containing the terminal verifier-level cause, the causal status of the relevant agent behavior, and the abstract agent mechanism exposed by the trace, to uncover the root causes. Harness proposal: propos…
    Harness Engineering for Self-Improvement | Lil'Log
  • Oddish is my attempt to build the tooling this loop needs: reliable trials, trace capture, provider-aware scheduling, and enough structure for agents to keep improving environments before human review. The full code is open-source here under the Apache 2.0 License.
    RL environment creation is becoming continuous QA - Rishi Desai
  • Meta-Harness (Lee et al. 2026) moves another level deeper: the optimized object is the code that determines and optimizes what information should be stored, retrieved, and presented to the model. “Meta-” in its name means it is a harness for optimizing harnesses.
    Harness Engineering for Self-Improvement | Lil'Log
  • The proposer for creating a new harness is itself a coding agent and the final output is a collection of harness candidates on the Pareto frontier. The entire execution history is accessible via a file system, and thus the coding agent uses commands like grep or cat to read through it instead of shoveling everything into a single prompt context. The proposed harness is a dictionary in the file system containing its own source code, scores, rollout trajectories, and state updates. The mete-harness loop iteratively creates new harnesses, and only qualified ones are kept.
    Harness Engineering for Self-Improvement | Lil'Log
  • The bi-level optimization is to find the best context given skill on the training data, while the outer loop finds the optimal skill that provides the best performance on the validation set: The skill database tracks the history of previous skills, context functions and eval metrics . A meta-level agent performs agentic crossover over prior skills to create a new skill given a task : . Then a base-level context engineer executes the skill and learns the context function from rollout feedback , guided by the current skill: .
    Harness Engineering for Self-Improvement | Lil'Log
  • The fact that ACE learns insights from rollouts helps us move toward self-managed memory, but the update rules and the overall workflow are still handcrafted. To move toward a more self-improving loop, Meta Context Engineering (MCE; Ye et al. 2026) separates the mechanism (how to manage context) from the artifact content (what is in context), running skill evolution at the meta-optimization level and context optimization at the base level. An MCE skill defines a context function and maps an input to context , where: are static components (prompts, knowledge bases, code libraries). are dynamic …
    Harness Engineering for Self-Improvement | Lil'Log
  • Generator: produces task trajectories, with reference to bullet points. Reflector: distills insights from successful and failed trajectories. Curator: updates the structured context with incremental, itemized entries.
    Harness Engineering for Self-Improvement | Lil'Log
  • Agentic Context Engineering (ACE; Zhang et al. 2025) treats context as an evolving playbook rather than an increasingly lengthening prompt. It has three components to maintain one context playbook of bullet points, each with an identifier and a description.
    Harness Engineering for Self-Improvement | Lil'Log
  • I(G;R∣history)≤I((S,Adv);R∣history)=I(Adv;R∣S,history)≤H(Adv).
    LoRA Without Regret - Thinking Machines Lab
  • The claim of 1-bit-per-episode may only apply narrowly to policy gradient algorithms.
    LoRA Without Regret - Thinking Machines Lab
  • For RL, we claimed that policy gradient algorithms learn roughly 1 bit of information per episode, given that there’s a single reward value at the end of the episode.
    LoRA Without Regret - Thinking Machines Lab
  • LLM datasets usually have a loss of around 1 bit (0.69 nats) per token, depending on dataset and model size.
    LoRA Without Regret - Thinking Machines Lab
  • has shown that neural networks can store 2 bits per parameter. These results pertain to the maximum amount of information absorbed in the long-training limit, not to the compute efficiency or rate of learning.
    LoRA Without Regret - Thinking Machines Lab
  • We can do on-policy training via reinforcement learning, by grading each student rollout on whether it solves the question. This grading can be done by a human, or by a “teacher” model that reliably gets the correct answer.
    On-Policy Distillation - Thinking Machines Lab
  • Distillation from large model teachers has proven effective in training small models to follow instructions,Alpaca: A Strong, Replicable Instruction-Following Model (Taori et al, 2021) reason on math and science,OpenThoughts: Data Recipes for Reasoning Models (Guha et al, 2025) extract clinical informationDistilling Large Language Models for Efficient Clinical Information Extraction (Vedula et al, 2025) from medical notes, and engage in multi-turn chat dialogues.Enhancing Chat Language Models by Scaling High-quality Instructional Conversations (Ding et al, 2023) The distillation datasets used …
    On-Policy Distillation - Thinking Machines Lab
  • more complex RL and OPD training workloads. In the post-training process of GLM-5.2, we used the slime framework to conduct parallel OPD training, efficiently merging more than ten expert models into the final model. The entire OPD training process took approximately two days, demonstrating high training efficiency.
    GLM-5.2: Built for Long-Horizon Tasks
  • As GLM-5.2 extends the maximum context length from 200K to 1M tokens, coding workloads are expected to shift substantially toward longer prompts. This shifts the primary inference bottleneck from computation to KV-cache capacity, long-context kernel overhead, and CPU-side overhead. Although the new GLM-5.2 architecture reduces per-token computational FLOPs, it does not proportionally reduce per-token KV-cache size. As a result, supporting longer contexts, higher concurrency, and higher token throughput under limited GPU resources becomes a central challenge for inference engine optimization.
    GLM-5.2: Built for Long-Horizon Tasks
  • We improve the MTP layer of GLM-5.2 for speculative decoding with two objectives: 1) Minimize the cost of the MTP layer as draft model; 2) Maximize the acceptance rate of speculative decoding. For the first objective, we also apply IndexShare on the mtp layer. In multi-step MTP, the indexer is placed on the first step and topk indices are used for all the following steps. However, different from the backbone, the input tokens of different mtp steps are different. As the following figure shows, if we reuse the topk indices of ℎ 4 h 4 ​ for ℎ 5 h 5 ​ , ℎ 5 h 5 ​ can only attend to ℎ 1 h 1 ​ to ℎ…
    GLM-5.2: Built for Long-Horizon Tasks
  • Contextually Private Mechanisms
    Zoe Hitzig
  • RL for Long-Horizon Tasks. For GLM-5.2, long-horizon tasks produce substantially longer execution traces, and once a super-long trajectory is split by compaction into multiple sub-traces, different rollouts under the same prompt yield different numbers of trainable traces with highly variable lengths. We therefore move from group-wise optimization to a critic-based PPO formulation that learns from individual rollouts, relying on a critic to estimate token-level advantages rather than group-relative comparisons. This single-rollout formulation fits compaction naturally, as it places no constrai…
    GLM-5.2: Built for Long-Horizon Tasks
  • The companies that manage AI costs best build a culture where efficiency is treated as an engineering achievement, not a constraint. Teams celebrate getting the same outcome with a smaller model, less reasoning, or a slower tier just as much as they celebrate shipping the feature itself.
    You're Spending Too Much on AI. You're Also Using Too Little. — Ramp Builders Blog
  • Smart defaults prevent unwanted expenditure from people who never touch their settings. But defaults can only go so far. Once someone starts making explicit choices, they are almost always incentivized to spend more.
    You're Spending Too Much on AI. You're Also Using Too Little. — Ramp Builders Blog
  • Firms are getting smaller. You cannot all work at Google. The layoffs are starting and they’re not going to stop. (Sure, most weren’t caused by AI, but agents may be the reason they don’t backfill roles.) Get good at leveraging AI in the real world—go home and find a mom-and-pop shop to make more efficient. Put agentic workflows in places they don’t belong. Start a business selling complements to AI: host conferences, deploy robots, collect esoteric datasets. Don’t bet on companies that only do desk work. Run the most efficient human services business in the world.
    The Old World Is Dying: Advice for 2026 graduates
  • If you are young, get ambitious. The traditional prestige paths are closing anyways. Start companies trying to design tech that will keep humans economically relevant and spread abundance.
    History is Yours to Write - The Intelligence Curse
  • Others are ignoring the AI capabilities in front of them, writing manifestos about what AIs cannot do followed shortly thereafter by proof that they can. Rather than changing their beliefs, they ignore the evidence.
    History is Yours to Write - The Intelligence Curse
  • Instead of staying at your job too long, take small symbols of “oh no I don’t feel appreciated here” and leave. Take advantage of small moments of wonder[3] during your life, seeing a deer race by, and google what breeds of deer live in your area. If a friend asks you to hang out, instead of saying no… why not flip a coin? By reading into the symbols around you, you can slowly start to build pseudo-structure to take more actions that you care about.
    the art of reading tea leaves
  • Theres also a second group of people, that don’t spend a lot of time thinking about what they do. I think this is not great (spending 80% of your time not thinking seems bad). There is a way to rephrase this internally though as “what are you doing” which could be so many things—playing deltarune, going on hikes, enjoying life etc. Some variations of initial questions I tend to ask people (please tell me more): What do you do? What have you been thinking about lately? (A good less direct question) Maybe a good riddle or fun problem I’ve been thinking about. (email me with your favorite ones!) …
    in defense of what do you do?
  • We were basically right about the price and advantages this offered, and did not substantially overestimate the amount of time / effort it would take. While the improvements list is longer than this, most of those are minor; fundamentally we built a cluster rivaling massive clouds for 40x cheaper.
    How to Rack 30 Petabytes of Storage | blog
  • Our software is 200 lines of Rust code for writing (to determine the drive to write data onto) and a nginx webserver for reading data, with a simple SQLite db for tracking metadata like which heap node each file is on and what data split it belongs to. We kept this obsessively simple instead of using MinIO or Ceph because we didn’t need any of the features they provided; it’s much, much simpler to debug a 200-line program than to debug Ceph, and we weren’t worried about redundancy or sharding. All our drives were formatted with XFS.
    How to Rack 30 Petabytes of Storage | blog
  • We cared a lot about getting this built fast, because this kind of project can easily stretch on for months if not careful. Hence Storage Stacking Saturday, or S3. We threw a hard drive stacking party in downtown SF and got our friends to come, offering food and custom-engraved hard drives to all who helped. The hard drive stacking started at 6am and continued for 36 hours (with a break to sleep), and by the end of that time we had 30 PB of functioning hardware racked and wired up. We brought in contractors for additional help and professional installation later on in the event.
    How to Rack 30 Petabytes of Storage | blog
  • London and Paris have surprisingly few successful tech startups for their size. Stockholm, a city of less than 1 million people, has Spotify, King, Klarna, iZettle, and Mojang, all valued at more than $1 billion. What's true of Stockholm that isn't true of other European cities? (Similar questions apply to Provo, Utah, and Tallinn, Estonia.)
    Questions · Patrick Collison
  • As a consequence, we need ever more domain-specific software. Rather than use universal tools for handling charts and for manipulating data, we tend to use separate analytics packages for every conceivable application. This is not all bad. Domain-specific tools can maximize ease-of-use and help amortize the cost of complex, specialized functionality. Sublime's built-in ⌘-T works better than every third-party Emacs package. Still, despite these benefits, the popularity of macros and browser plugins strongly suggest that users are smart and want more control.
    Questions · Patrick Collison
  • Whatever is going on, a growing fraction (about 55% today) of a growing number of humans live in cities. Given that our physical location determines so much of our quality of life, shouldn't we be very distressed that we seem to be getting worse at making them? And what else are we getting worse at?
    Questions · Patrick Collison
  • On policy matters, what determines when someone preferences other values ahead of their personal benefit?
    Questions · Patrick Collison
  • The early days of the US Interstate Highway System saw a construction cost of about $1.5M per mile in 2019 dollars. (Source.) San Francisco currently projects that a bus lane will cost $309.3 million over 2.2 miles, i.e. $140M per mile. This isn't quite like-for-like (SF is merely making some improvements in a readily-accessed urban environment, not constructing de novo highway), but is nonetheless paying 93x more per mile.
    Questions · Patrick Collison
  • When New York decided to build a subway, the first contract was awarded on February 21, 1900. 28 stations opened and general operation commenced on October 27, 1904, 4.7 years later. The total cost was about $1.1 billion in 2019 dollars. (Source.) In April 2000, the MTA decided to build the Second Avenue Subway. The first phase, with 3 stations, opened on January 1, 2017; it cost $4.45 billion. Despite construction technology having presumably improved in the intervening century, the Second Avenue line was 37x more expensive on a per-station basis.
    Questions · Patrick Collison
  • Agent identity. Claude Cowork's answer to agent identity is concrete: credentials stay in the host keychain, the VM gets a per-session scoped-down token, and that token can be revoked independently of the user's. However, we are starting to grapple with the broader question of cross-platform agent identity. Should an agent possess its own principal identity, or should it act as an extension of the user and inherit the user’s permissions? Ultimately, the answer may be a blend of the two.
    How we contain Claude across products \ Anthropic
  • Multi-agent trust escalation. On the one hand, sub-agents can isolate untrusted content, returning structured facts rather than raw text up to the main agent. On the other hand, this can be abused: if a sub-agent's output is treated as higher-trust than raw tool results, because such output came from "us,” a new vector for prompt injection is introduced. In multi-agent systems, there is a tradeoff between allocating differing trust levels and becoming liable to trust escalation.
    How we contain Claude across products \ Anthropic
  • Persistent memory poisoning. The share of agent context that persists across sessions keeps growing—this includes product memory, CLAUDE.md files, mounted workspaces, and the state directories of scheduled and long-running agents. An injection that lands in any of these is reloaded each time the agent starts. As more agent state survives the session, we are threatened by new persistence mechanisms in the classic post-exploitation sense. Good classifiers on session startup will need to become more commonplace.
    How we contain Claude across products \ Anthropic
  • Risk we missed: VM isolation kept the endpoint detection software out too
    How we contain Claude across products \ Anthropic
  • The fix in each case had the same shape: defer parsing and execution of project-local configuration until after the user accepts the trust prompt. If you're building something similar, treat project-open, config-load, and localhost listeners the way you'd treat any inbound request from the internet. They shouldn’t be implicitly trusted just because they feel local and arrive before the user has consented.
    How we contain Claude across products \ Anthropic
  • The external content the agent can reach. MCP servers, third-party plugins, and web search tools all feed content into the agent’s context from sources you don’t control. An audited connector isn’t the same as audited data—a GitHub connector, for instance, can load a poisoned README straight into the model’s context despite passing malware checks. Granularly limiting tool permissions can help limit the blast radius. An agent with read-only DB access, for instance, can be deployed far more broadly than one that writes to prod.
    How we contain Claude across products \ Anthropic