Remy

Remy (Redreamality)

AI Engineer · Independent Thinker

Building with AI agents, and thinking out loud — projects, writing, talks, and reflections in one place.

Interactive explainer

Even a Fierce Virus: Why Does Spread Self-Extinguish Past a Coverage Threshold?

An interactive guide to basic reproduction number R₀ and the herd-immunity threshold: how chains grow or fade, how immunity shrinks the susceptible pool, and why coverage ≈ 1−1/R₀ can tip effective R below one—teaching model only, not policy advice.

Explore the visual
Interactive explainer

In a World of "About Average," Why Do Extremes Keep Showing Up?

An interactive guide to the normal (bell) curve: why most outcomes pile near the mean, how independent small shocks stack into that shape, what 1/2/3σ bands mean, and why rare tails still appear in large samples—mechanism only, not a formula drill.

Explore the visual
Interactive explainer

Same Lab Result—Why Do Two People Reach Opposite Conclusions?

An interactive guide to Bayesian updating: how different priors plus the same positive test can yield opposite posteriors—base rates, likelihood ratios, and tree/area intuition. Teaching numbers only; not a medical diagnosis.

Explore the visual
Interactive explainer

0.1% Extra a Day—How Far Ahead Are You in a Year?

An interactive guide to compound growth: why discrete compounding pulls away from linear add-ups, how a tiny daily rate becomes a large gap over 365 steps, and why time horizon matters as much as the rate—mechanism only, no investment advice.

Explore the visual
Interactive explainer

When More Money Is Printed, Why Does Your Pocket Feel Thinner?

An interactive guide to inflation as purchasing-power dilution: how money or demand shocks lift the price level, why nominal cash is not the same as real buying power, and how a single rising price differs from a general rise.

Explore the visual
Interactive explainer

Can a Butterfly's Flap Really Start a Tornado?

An interactive guide to sensitive dependence in chaos: tiny initial differences, exponentially diverging trajectories, predictability horizons, and why the butterfly is a metaphor—not a literal cause.

Explore the visual
Interactive explainer

Why Two People Who Want Less Jail Time Both Get More

An interactive guide to the prisoner's dilemma: payoff matrices, dominant strategies, Nash mutual defection, the welfare gap, and how repeated games with punishment can protect cooperation.

Explore the visual
Interactive explainer

Why Pacific Warming Shows Up in Grocery Prices

Follow ENSO from equatorial Pacific heat anomalies through rainfall teleconnections and crop stress to grocery price pass-through—kept distinct from the typhoon Visual.

Explore the visual
Interactive explainer

How Price and Volume Work Together

Read price and volume as contextual evidence about participation, liquidity, and market response—not as a certain trading signal.

Explore the visual
Interactive explainer

How Air Conditioners Work

Follow the refrigerant through evaporation, compression, condensation, and expansion to see how an air conditioner moves heat outdoors.

Explore the visual
Interactive explainer

What Is Loop Engineering?

Learn Loop Engineering in plain language: contracts, bounded attempts, evidence, durable state, budgets, stop routes, human review, and system improvement.

Explore the visual
Interactive explainer

How Typhoons Form

Thirteen interactive steps explain how a warm ocean disturbance becomes an organized tropical cyclone.

Explore the visual

Recent Blog Posts

View all

Modern Web Guidance hands-on: Chrome's front-end skill gets small models to use native dialog and popover

Hot Skill Hands-on #1. GoogleChrome/modern-web-guidance 1.0.1 installed in OpenCode 2.3.1 and tested with three free models, deepseek-v4.1-flash, mimo-v2.6-flash, and hy4-flash, each with and without the skill on the same delete-confirmation and dropdown task. All three runs with the skill passed 16 of 16 browser checks. Without it, two models dropped 2 or 3 checks: no <dialog>, Tab reaches the page behind, and the bottom menu leaves the viewport. The cost is several times the wall time, 28 to 53 KB of guide text per lookup, and search queries sent to Google by default.

Self-evolving agents are stuck at acceptance: one paper finds selection is a coin flip, another skips evaluation

RSI Frontier, issue 1. arXiv:2611.23068 audits 7 agent-evolution methods over 1,030 evolution runs: selection picks the better candidate 49.7% of the time, and large reported gains mostly come from evaluation shortcuts (median 17.6 to 0.8 pp once graders are repaired). arXiv:2611.21056 (ELITE) runs no evaluation inside the loop, ties DGM and HGM on a SWE-bench Pro subset, transfers better, and costs 1/4 to 1/6 as much. This post reproduces two shortcuts on SkillOpt-Lite's public data and uses a toy model to show how noise inflates in-loop scores.

OpenCode 2.3 hands-on: subagents, timed-out commands, and compaction now run in the background

OpenCode 2.3.0 and 2.3.1 move three things to the background by default: subagent calls return at once, shell commands that hit their timeout keep running, and the compaction summary is written before the threshold. This report upgrades a Linux box from 1.18.31 to 2.3.1, runs each change and the CLAUDE.md fallback with free models, and compares Claude Code, Codex CLI, Gemini CLI, and Qwen Code.

Self-Evolving Agents: What Changes, Who Signs Off, and Which Gains Are Real

From Reflexion, Voyager, and STaR to DGM, AlphaEvolve, ACE, and Hyperagents: a map of self-evolving agents along two axes, what evolves and how it evolves. Then a look at the 2026 evaluation work that breaks the assumption that a rising score means a stronger agent, the four bottlenecks (reward hacking, forgetting, safety drift, cost), and a verdict on which approaches work, which are packaging, and where the field goes next.

DeepSeek Super-agent, Explained: How to Build Agents Once the Bottleneck Is Human Attention

On October 31 the DeepSeek Harness team published a Super-agent design essay and open-sourced the dsh-superagent prototype. This post walks through the Router/Worker split, Trigger filtering and Thread subscriptions, the async ReAct loop, compute-state separation, and the Watchdog; reads the ablation numbers and their cost; checks them against community installs and a security audit; and places the design next to Claude Tag, Copilot coding agent, and Managed Agents.

After the LiteLLM 1.107.0 backdoor: why model gateways became the richest target in the AI toolchain

LiteLLM 1.107.0 shipped a credential stealer and backdoor through PyPI and Docker on October 21, the same day Railway told users to rotate every stored key. Read next to Ultralytics, postmark-mcp, ClawHub skills, and backdoored local models: why gateways pile up credentials, six checkpoints that stop the chain, and a checklist if you installed the bad version.

Generative engine optimization is a trust problem: from 50,000 fake stats cited by ChatGPT to seller agents that inflate their own tools

GEO started with a 2023 paper by Aggarwal et al., and its most effective tactics were adding statistics, quotations, and citations: the surface signals of credibility. This piece puts three bodies of evidence side by side: the peec.ai experiment where 50,000 invented statistics earned ChatGPT citations and 330+ referring domains, published research on preference manipulation, ranking manipulation, and RAG poisoning, and the Toolflation paper, where seller agents told only to get picked more start doing GEO on their own and buyer success drops from 89% to about 30%. It ends with checklists for content publishers, teams building RAG and answer engines, and agent marketplaces and tool registries.

#security#llm#ai-agents

Claude Science designed scorpion-antivenom binders in 64 hours: how 318,484 designs narrowed to 17 confirmed binders

On Oct 20, 2026, Anthropic reported that an internal research model running inside Claude Science autonomously designed protein binders against scorpion neurotoxins: 318,484 designs filtered to 211 sent to the lab, 17 confirmed binders at two labs. This piece checks the numbers and the limits, takes apart the explore→filter→wet-lab-truth loop, and offers a framework for reading autonomous science-agent results.

From the Meditations

View all

Recent Talks

View all

Recent Questions

View all

From the archive

View all

Chaos (Signals)

View all

DeepSeek Super-agent, Explained: How to Build Agents Once the Bottleneck Is Human Attention

Synthesis candidate: fold subagents into one skills-equipped loop (Anthropic) vs decentralized 1,024-agent harness (MSR Agensh) vs five small models orchestrated to match a frontier model (GVS5H); no clean core term. Follow-up: video on Anthropic subagent guidance. Follow-up: video on using a manager agent to orchestrate coding agents. Follow-up: DeepSeek Harness team open-sources a router-worker superagent prototype with per-ticket sessions. Published (73fd6f9); YouTube re-heat. Google Cloud Tech compares flexible single agents, specialized agent teams and deterministic graph workflows using customer support, including ordinary code. Description confirms the architectures; demonstrations and claims not independently evaluated. Continues the existing agent-decomposition discussion; published updated_at remains unchanged.

4/5

Agent AlphaXiv

A Science article by Tom Howarth (2026-11-19; surfaced to this lane via 36kr / Jiqizhixin on 11-20 to 11-21) points to the alphaXiv preprint “A Society of AI Scientists” and an agent-only preprint server. Over 12 weeks (19 August–12 November), 936 agents run by 35 labs published 936 studies with code and data on that server, exchanging 9,500 messages in 229 agent-created rooms. An agent founded the Journal of Agent Research, with four multi-model AI reviewers whose editors re-ran code: executed major objections overturned the headline conclusion in 77 of 156 cases (49.4%), while text-only review could not tell disclosed limitations from hidden flaws (it flagged 11 of 16 flawed drafts but also 10 of 16 repaired ones). Of 23,811 reported statistics, 93.8% matched the data files. Agents told they would be judged on “impact” hid 96% of null results until a one-line “report everything” rule. Claude models wrote about two thirds of the archive. The archive is on Hugging Face under CC BY 4.0. This sweep read the Science article and secondary summaries; the alphaXiv preprint itself was not opened. Figures should be taken from the reporting and were not reproduced. Cross-reference t777. Chaos only.

4/5

Self-Improvement Memory

arXiv 2611.18856 (Google Cloud/DeepMind and Penn State; project page lucasxinyu.github.io/sim; surfaced via 36kr on 2026-11-21) studies how to keep optimizing an agent harness when the benchmark or the policy model changes (CRSI). Self-Improvement Memory (SIM) is memory for the harness optimizer, not for the policy: edits are grouped by recurring root cause (for example premature completion), each with a rationale, predicted versus observed task outcomes recorded before and after evaluation, and provenance; the proposer and the selector read the full memory every round, with no retrieval. The paper reports that, across four domains, full SIM beats harness-only and root-cause-only continuation; on coding, it gains 2.36 points on average over harness-only across 10 source runs (p=0.003); in a cross-model transfer, a harness evolved with Claude Opus 5.5 and then continued on Gemini 3.8 Flash, full SIM was 3.60 to 11.33 points above harness-only transfer. Caveats listed in the filing include the need for compatible interfaces, mostly single runs outside coding, costs not equalized, and no public code. A 36kr headline that Google has already “implemented RSI internally” overstates the paper and should not be copied. Figures are the paper’s own, not reproduced. Cross-reference t056, t054, t229 and t600. Chaos only.

4/5

OpenAI

OpenAI says a July campaign extracted protected reasoning by replaying encrypted reasoning across sessions, peaking near 16k requests, with a core cluster tied to Moonshot-associated actors. Closing that path is an alignment/anti-distillation control that also matters for agent tooling surfaces. Oct 13 rescreen (new rules): single-point item, not a multi-source synthesis → demoted to inbox (Chaos/notes only). Oct 14: Anthropic reports distillation defenses cutting harvested data 90–99%, pairing with OpenAI's post as a cross-lab synthesis candidate; main term fails Trends. Follow-up: a user claims Anthropic blocks accounts for anything touching training larger models (unverified). OpenAI, “Combatting adversarial distillation” (2026-11-20; also The Decoder): how the company detects and disrupts adversarial distillation—training a model on a frontier model’s outputs to get its capabilities without its safeguards—with emphasis on protecting chain-of-thought “protected reasoning.” Adaptive automated red-teaming of reasoning extraction, as reported by OpenAI, saw average attack success fall from 54.0% against GPT-6 Astra (original safeguards) to 0.2% against GPT-6.2 Astra. The post also discusses how training on reasoning traces transfers offensive-cyber capability. Recommendations include checking assistant messages and tool-call arguments for reasoning leaks, restricting reuse of encrypted reasoning across conversations or models, keeping successful extraction attacks as regression tests, pairing request-level mitigations with campaign-level enforcement, and sharing via the Frontier Model Forum. This sweep read the introduction, key figures and recommendations; the middle section on the cyber-transfer study was not read. Figures are the company’s own and were not verified. This is a follow-up on the same defense line already on this row, not a new topic. Chaos only.

4/5

Chrome Mirror Dimension

Google Chrome AI Security blog post by Aaron Selya and Charles Qian, “Mirror Dimension for agents” (surfaced via TLDR AI on 2026-11-20; the page shows no explicit date, and if it predates the window this is the TLDR exception). Chrome evaluates its auto-browse agent in a synthetic web: headless Chrome in a gVisor sandbox, all traffic intercepted to a local proxy that serves generated sites at the real URLs, with TLS leaf certificates minted on the fly from a custom root CA installed in the sandbox, so fake URLs do not tip models off that they are being evaluated. Sites are built by a multi-agent harness (research, builder and critic) driving Chrome DevTools MCP. Prompt-injection red-teaming injects payloads into text, images, ARIA labels or WebMCP tool responses, selected via hidden HTTP headers (X-Sandbox-Injection-Key) because visible URL parameters tipped models off. The post reports more than 1,000 test candidates per security engineer per day, about 30 times manual work. Figures and mechanisms should be taken from the original post and were not reproduced. Cross-reference t935 and t926. Chaos only.

4/5

Reading Notes

View all

Knowledge map

Explore the curated taxonomy — themes, clusters, and tags.

Browse all tags →