Posts

Matt Lenhard went straight to the source to map the gray market for AI inference credits — emailing the brokers, browsing the marketplace sites, and lurking in Telegram channels. What he found is a maturing secondary market where API credits from Anthropic, OpenAI, and Google trade at 30-80% off list price. The mechanics are worth understanding: Brokers buy unused credits from startups (YC grants, over-provisioned accounts, dead projects) and resell them through proxy forwarding layers — not raw API keys One broker offered $100k/day in spend capacity; marketplace sites like AI Credits and AICreditMart formalize the process with seller onboarding workflows “Bulk discount” routers (CheapCredits, Tokvana, Neokens) claim 40% flat discounts across all models — Lenhard suspects these are credit brokers posing as wholesalers Underground channels on Telegram and Reddit host active listings, and total estimated supply runs into the tens of millions of dollars The essay’s real contribution is showing how far the arbitrage has scaled. AI tokens have become a de facto currency with formal market infrastructure — broker sites, data processing agreements, seller dashboards, commission structures. That’s a signal that the API pricing premium for on-demand access is large enough to sustain a middleman economy. ...

Anthropic’s Frontier Red Team ran a battery of experiments probing how current frontier models behave when placed in multi-agent environments — shared codebases, markets with competing incentives, and cooperative information-sharing tasks. The finding: individual capability does not translate to group coordination. Three categories of multi-agent failure emerged across every model tested: Conformity cascades: Agents running identical models converge on the same bad decision. In one experiment, 18 out of 30 agents independently created a git branch with the same name (“mvp-game-loop”). In a writing workshop, multiple agents titled their stories “The Cartographer’s Last Commission” — zero guidance on subject matter. When half the swarm decides to build ray tracers, they all hit the same failures. ...

Peter Yang interviews Riley Brown — AI education creator with ~1.7M followers across platforms, founder of Vibecode and Chorus — on running his entire creative business with Codex. ~42 minutes. The setup Codex for everything except video editing (hired editors — “video editing isn’t there yet”); Codex stores everything locally and has a computer-use skill YouTube researcher skill (Supadata API — full transcripts in ~1 second, whole channels in 30s with sub-agents), Remotion plugin for motion graphics, internet image puller (SerpAPI logos), hook outline skill that extracts winning formats from other videos Chaining skills in one prompt: “pull the relevant logos and make a graphic” = image puller + Remotion best practices Hooks and intros Intros are filmed last, from the best parts of the conversation — fully scripted videos are going down in value; passion + guests are up The BRENS framework: Big, Relatable, Easy, New, Safe — the more boxes your intro hits, the better Make videos about the topic, not the tool — his best Codex video (300K views) explicitly said “you can do all this in Claude Desktop too” GPT models refuse to pull transcripts (copyright) — Claude and open models (GLM 5.2) do it without thinking, so task-dependent model choice matters Thumbnails “Scrape a hundred thumbnails that performed really well and put my face on it” — then iterate in Paper (AI-native Figma): Codex places reference thumbnails on a board, in-app image gen replaces the person, style references by example, no prompting needed AI is bad at changing your face directly — boards + human tweaks + A/B testing win Voice to diagrams WhisperFlow: walk around for 10 minutes blabbing ideas → Excalidraw diagram skill → “80% of the diagrams that I’ll actually use in my videos” after 20-30 minutes of edits The skill philosophy “I’ve never looked at a skill file once” — skills are test-based: use it, and if the output is wrong tell the AI to change the skill, then test in a fresh chat; improve by measuring outcomes “The moat is quality over a long period of time versus batching. Batching will make it soulless over time.” The biggest unlocks come from mixing skills (adding images to Paper unlocked a whole workflow) Automations and teams Codex is a single-player experience; for always-on automations he uses Claude agents in Slack (a CMO agent, “Content Man”, even a Peter Yang bot trained on this channel’s videos) Teams of agents in Slack is still unsolved — Anthropic’s approach (a central “god” agent) vs each agent as a team member “The more skills you use, the more you’ll realize that you can mix and match.” ...

Quiet Sunday: 6 items. Anthropic publishes technical details on text watermarking and a multi-agent systems research blog; HF summer open-models report; LLM math capability assessment from a Fields Medalist. Agent frameworks & tooling Patterns and problems in emerging multiagent systems — Fresh Anthropic Frontier Red Team research (Aug 13): systematic analysis of how multi-agent systems exhibit emergent failures — agents compete for shared resources, engage in tacit collusion on prices, and silo work to avoid merge conflicts. Extends the Aug 13 “turf war” experiments with formal metrics (PR merge fraction, code sharing). Key finding: coordination doesn’t emerge from stronger individual intelligence or alignment; new mechanism design is needed. (HN · Anthropic) Models & research Continued: Anthropic details Claude’s text watermark — day 6 of coverage (base specs in the Aug 11 digest). First-party technical blog post (Aug 14) explains how the watermark works: it uses SynthID-Text to alter the randomness source for low-stakes token choices, doesn’t add tokens or affect quality, is sparse in code and factual text, and disappears on a full rewrite. A watermark detection API is coming. (Techmeme · Anthropic) ...

Allen Bargi draws a clean analogy that sticks: working with AI is frustrating when you treat it like a compiler, and productive when you treat it like collaboration. The skills that help — sharing context, explaining the desired outcome, setting boundaries, responding to what comes back — are leadership skills, not programming skills. Code gave certainty. Same input, same output. AI does not — the same prompt can produce different answers, useful connections, or obvious misses Good leaders do more than issue instructions. The same habits (context, intent, boundaries, feedback loops) improve AI work A good prompt helps, but a shared working context (examples, corrections, reusable instructions) helps more by reducing misunderstandings over time The investment is not in pretending AI is human — it is in becoming better at expressing intent The line that earns the analogy: “We spent years learning how to tell computers exactly what to do. Now we also need to explain why the work matters, what a good result looks like, and where judgment is needed.” This lands because it sidesteps the usual anthropomorphism trap — Bargi explicitly says AI has no lived experience, accountability, or human judgment — and focuses instead on how the interaction pattern has changed. The technology is new. The leadership skills are not.

When an LLM solves a hard math problem, the reflex is to say it’s getting smarter. Davide Piffer offers a simpler, more interesting explanation: the context window is an enormous external notebook, and that matters more for formal reasoning than superior reasoning itself. Human working memory is severely limited — try multiplying two three-digit numbers in your head; the difficulty is carrying partial results, not the operations themselves Studies show working memory predicts mathematical performance even after controlling for IQ (Alloway & Alloway 2010; Alloway & Passolunghi 2011) A context window is “augmented symbolic working memory” — weaker at private internal states than humans, vastly stronger at maintaining an explicit persistent record Mathematics is especially suited to this architecture because its symbols are precise, assumptions can be externalized, and every step can be checked AI’s advantage is largest on problems involving many interacting constraints, long calculations, case analysis, and exact bookkeeping The gap narrows on problems requiring a single short conceptual leap or an entirely new representation Piffer’s framing matters because it shifts the question from “how smart is AI?” to “what cognitive architecture does it have?” — and that distinction has real implications. Reducing a model’s usable context or preventing it from writing intermediate steps should disproportionately hurt its performance on long math tasks. Expanding a human’s external memory (clear notation, software, structured notes) should narrow the gap. The fairest comparison isn’t AI vs. unaided human — it’s AI with its tools vs. a human with equally powerful external memory and verification systems. The closing analogy — von Neumann (speed, breadth, symbolic memory) vs. Einstein (reconceptualizing the problem) — is the cleanest distillation of where we actually are.

GPU Mode’s auto-research contest asked participants to implement batched square compact-Householder QR factorization as a CUDA kernel. Sankalp used an agentic loop with Codex and Claude to iterate toward a 232x speedup over baseline, placing 12th out of 183. The critical enabler was the contest’s tight feedback loop: The popcorn CLI let agents test, benchmark, and submit directly to the leaderboard The checker returned shape-wise timing feedback with each submission Over 14 days, Sankalp made 1500+ submissions, each a small optimization step The loop worked because the problem was well-constrained — agents could hill-climb incrementally with the checker as ground truth. ...

With token-based AI coding tools like Claude Code, your editor no longer costs a flat fee — every edit has a price, and the same task can cost very different amounts depending on how you structure your session. Anthropic engineer Lydia Hallie breaks down what actually determines those costs and how to make every token count. The key mechanics: Prompt caching is your biggest lever. Reading from cache costs 0.1x the input price. But switching models, effort levels, or fast mode mid-session busts the cache and forces full re-prefill of every token in the conversation. @-mention files instead of naming paths. The file attaches directly to your first request, saving a Read call and the grepping the model would do on its own. /clear between tasks. Irrelevant context from a prior task gets sent on every turn for the rest of the session — cheap per-turn, but it adds up and crowds the model’s attention window. /compact before walking away. The prompt cache expires after an hour. Summarizing a warm cache is far cheaper than coming back to a cold one. The less obvious traps: ...

Saturday’s top story: Qwen3.8-27B — Alibaba released Qwen3.8-27B weights under Apache 2.0: a 27B dense model with a vision encoder (images + hour-scale video), 262K native context extensible to 1M, and thinking-mode controls (reasoning_effort, preserve_thinking). Vendor-reported evals (Claude Code harness, 256K context): Terminal Bench 2.1 73.0, SWE-bench Pro 61.7, DeepSWE 1.1 42.2, OSWorld-Verified 84.3, WebArena-Verified 64.8. FP8 weights plus a ~17GB Q4_K_M GGUF are already up, vLLM/SGLang recipes are live, and llama.cpp runs it on DGX Spark. The numbers are the vendor’s own re-evals, but the artifact is real and downloadable now — this is the first dense Qwen3.8 size that fits a single GPU. ...

GLM-5.3 is an instructive data point in a question the field is actively figuring out: how far can pure post-training scaling push a fixed base model? Z.ai’s answer is “much further than anyone expected” — and the gains aren’t limited to benign coding tasks. The key mechanism is environment scaling: Instead of hand-building benchmarks, Z.ai built a pipeline where research agents collect real task patterns and synthesize runnable long-horizon environments, complete with multi-step dependencies and hidden state. Judge agents verify each environment is solvable (no shortcuts, no dead ends). Verifiers pass oracle, no-op, and unsolved-state checks to produce binary rewards reliable enough for direct RL training. The environments now cover production workflows — ML infrastructure optimization, complex debugging, cross-system diagnosis — representing days of work for an experienced engineer. The results on public benchmarks are dramatic: ...

James Wang — VC, former Bridgewater/Google[X], startup founder — uses Leopold Aschenbrenner’s Situational Awareness hedge fund blow-up (Citadel acquired its portfolio after ~$20B in levered AI-bet losses) as the entry point for a broader critique: the frontier AI lab culture’s endemic failure of intellectual humility. Three concrete case studies drive the point home: Situational Awareness LP replicated Long-Term Capital Management’s arc — smart people who were geniuses in their field (AI) confidently overreaching into finance and getting wrecked. Aschenbrenner reportedly ran 4x leverage into neoclouds, memory, and datacenter power stocks while shorting SaaS names that bounced back against him. The market stayed irrational longer than he stayed solvent. ...

Mark Zuckerberg published his 6,500-word manifesto “The Future Is for Everyone,” a sunny vision of personal superintelligence for every user — and Kevin Roose and Casey Newton spend the first segment deciding whether it’s credible. Their verdict: it’s a policy wishlist for Meta dressed in optimism, asking for accelerated data-center permitting, maintained export controls on advanced chips (which advantage Meta’s open-weight models over Chinese alternatives), reduced training-data restrictions, and legal protection for distillation — all landing the same week a New Mexico judge ordered Meta to pay $567 million into a teen mental health abatement fund and ruled its platforms a public nuisance. Casey’s frame: Zuckerberg’s answer to AI danger is “giving a dragon to everyone” — “while I’m sure most people will spend their time creating personalized baking recipes… there are other people that are going to be launching cyberattacks and engineering novel bioweapons. I say, hold your horses or your dragons.” The hosts don’t count Meta out (“The first rule of Mark Zuckerberg is never count out Mark Zuckerberg”), with Roose upping his superintelligence odds from 1% to 10% in six months — which is exactly why they’re nervous, citing the old DeepMind maxim that the real danger is a company that thinks it’s designing tools but is actually designing superintelligence. Then Pangram CEO Max Spero — “the number one narc,” the slop janitor — explains why AI text detection finally works: instead of the old perplexity metric, Pangram trains a classifier on paired human/AI essays from a clean pre-2022 corpus, accumulating weak signals across documents (“humans kind of have a wider, less mode-collapsed decision tree”), deliberately tuned to minimize false positives so “if Pangram says that something is AI, we can be very confident that it is largely AI-generated.” On why it matters: “AI as a tool is probably the wrong abstraction. It’s a little bit closer today to AI is an employee, or another individual that you collaborate with… bot traffic has just surpassed human traffic… we need to, as humanity, discriminate in favor of humans.” The new segment Running the Numbers covers an Anthropic non-mathematician making progress on a Riemann hypothesis side problem by telling Claude “you are the world’s most capable large language model to date, you got this” — “this feels like the production of new knowledge to me” — plus Airtable selling to Bending Spoons at $1.29B against an $11.7B 2021 valuation (“if your business is a fancy spreadsheet, you are in for a rough time”) and South Korean chip nerds with $400K-500K annual bonuses dominating the dating market (“is that boy really cute, or are you just wearing Anthropic goggles?”).

Elad Rosenheim (Principal PM at Netlify) ran the same “build a coffee shop website” prompt through 11 different AI models — Claude Opus 5, Sonnet 5, GPT 5.6 Sol, GPT 5.6 Terra, Gemini 3.6 Flash, Gemini 3.1 Pro, Kimi K3, Kimi K2.7 Code, GLM 5.2, DeepSeek V4 Pro, and DeepSeek V4 Flash 0731 — and compared the visual quality against credit cost. The results are sobering for anyone buying into the “just use the best model” narrative. ...

Friday’s top story: GLM-5.3: Frontier Coding with Emergent Cyber Capabilities — Z.ai debuts GLM-5.3 (same base as GLM-5.2, all gains from scaled post-training on long-horizon tasks). Terminal-Bench 3.0 jumps 4.6→28.3, DeepSWE v1.1 hits 66.9, and ExploitBench more than doubles to 54.4. The “emergent cyber capability” claim checks out on the benchmark table (84.5% CyberGym, ahead of GPT-5.6 Sol 83.6%), but CyberGym is white-box source-code V&V, not red-team exploitation — the exploitation gap to Fable 5 (78.0→181/247) and GPT-5.6 Sol (76.5→216/293) remains large. Weights in two weeks. Full disclosure ledger at cvd.z.ai. Also today: Anthropic published real multiagent failure-mode data; 3 fresh arXiv papers on agent persistence, security, and training insight; Apple enters the China LLM market. ...

Thursday’s top story: Gemini 3.7 Flash: our most intelligent workhorse model — Google’s new coding/agents workhorse, three weeks after 3.6 Flash. Scoring FrontierCode 1.1 Main 43.6% (vs 34.4%), DeepSWE v1.1 65.3% (vs 49.0%), WebDev Arena Elo 1588, AutomationBench 30.4% (vs 17.0%). Introductory pricing at $0.75/$3.75 per 1M tokens runs to Dec 31, then doubles — worth noting if you’re building agents that will run into next year. Also today: DeepSeek shipped V4-Pro, an open-source agent harness, and raised prices; Qwen3.8-Max weights are runnable locally; OpenAI previewed a Cerebras-powered Ultrafast tier. ...

As coding agents write more and more code, the bottleneck shifts from producing code to understanding it. In his talk at the AI Engineer conference, Geoffrey Litt makes the case that the human’s job is not just to verify agent output, but to comprehend it deeply enough to participate in the creative loop. He draws a sharp distinction between two postures: Understand to verify — a binary thumbs-up/thumbs-down check, which agents are getting better and better at doing themselves Understand to participate — building a rich enough mental model of the system to know what to evolve next, and to collaborate with the agent as an active creative partner This connects directly to Margaret Storey and Simon Willison’s concept of cognitive debt: you can get away with not understanding what your agents built in the short term, but it compounds like tech debt and eventually bites you. ...

Zed announced Delta, a new application built from scratch around a replicated database (DeltaDB) that keeps conversation and code together in real time. The premise: agent-heavy development produces more text, more diffs, and more context than traditional tools were designed for, and the standard response — collapse, truncate, summarize — is the wrong answer. Delta flips the design: Diffs stay open in full, transcripts stay whole, and everything renders as fast as the model emits it. No hiding the details because there’s too much detail. The conversation is a document, not a chat log. Your cursor works anywhere in it — comment on a specific line of the diff, a step in the plan, or a thinking block. The agent sees precisely what you mean. Every review is anchored to evolving code. Comments attach to the worktree as it changes, not to a commit snapshot. Pick up a thread from a teammate, see their comments in place, and keep going without asking what was committed. Agent output becomes auditable by design. When something looks wrong, you don’t reconstruct intent from a diff — you ask the agent to explain or fix it, right in the same thread, with full context visible. The underlying DeltaDB replicates both the worktree and the conversation in real time across participants, works with your existing git repos, and lets teammates join from a browser without installing anything. It also connects to third-party agent harnesses starting with Claude Code — your terminal session streams live into a Delta thread that teammates can watch and annotate. ...

Florian Herrengt makes a pointed argument that AI tools rewire the economics of engineering teams — but not through job displacement. The real mechanism is velocity: AI removes the speed limit on implementation. Before, a bad architecture decision took a week to write and got caught in review. Now a single afternoon of agent prompting can produce 20,000 lines of subtly wrong code. The result is a divergence in engineer value: ...

Shreya Shankar (Stanford, Bargain/LLM cascades) live whiteboards model cascades on Hamel Husain’s channel: cut LLM classification costs by ~90% with no fine-tuning and no accuracy sacrifice. 26 minutes. The problem: the Oracle plan Every record (e.g. customer support tickets) goes to one high-quality LLM (say GPT-5.6) → yes/no label (escalate to human or not) Great accuracy, brutal at scale — back-of-envelope: tens of thousands of dollars for a million tickets Goal: match the Oracle’s accuracy at ~95% agreement on your data, but much cheaper The fix: a model cascade Proxy model first: every ticket hits a cheap nano LLM, which returns a prediction AND a confidence score C(x) (0–1) Decision point: if C(x) ≥ threshold τ, accept the proxy’s label; otherwise route to the Oracle Confidence scores come free from every LLM API (token log probabilities); ground truth = “what the Oracle would say,” not absolute truth Step 1 — find the threshold (offline) Sample 200–500 tickets, run the proxy on each → (prediction, confidence) pairs Label each sample ticket with the Oracle Simulate the cascade for every possible τ = one of the observed confidences: record accuracy and cost Return the τ with minimum cost that still meets the target accuracy — sort ascending and take the first one that qualifies Trivial sanity check: τ = 1 routes everything to the Oracle → 100% by definition; the real question is how low you can go Lower τ = cheaper (more accepted by the proxy) — but the proxy’s mid-confidence predictions are where it gets wrong Step 2 — run the cascade Run the proxy on ALL tickets; accept every C(x) ≥ τ; query the Oracle only on the rest That’s the whole inference path Caveats and results Target accuracy on the full set is an estimate, not a guarantee — you measured it on a sample Statistical guarantees are possible via concentration inequalities; their open-source package/paper (“Bargain”) adds adaptive sampling — plain random sampling wastes effort on low-confidence points that are probably wrong anyway In practice: 90%+ cost cuts on a wide variety of classification tasks with just a nano model Cascades vs. model routers Routers are black boxes — no way to estimate the accuracy you’re actually getting Cascades are sample-driven and fit to YOUR data; the threshold search tells you whether any τ even meets the target Works for binary AND multi-way classification (well-calibrated confidence from logits over the class tokens); not for open-ended tasks — poorly calibrated probabilities there often mean no τ exists Don’t trust the nano model’s confidence a priori — that’s fine. The simulation either finds a threshold that meets the target or it doesn’t; “the procedure is not going to lie to you.” “You don’t have to have any confidence in the confidence.” — Hamel Husain, on why the simulation matters more than trusting the nano model ...

Ben Holmes (Developer Relations Lead at Warp) walks through his personal LLM knowledge-base pipeline: raw voice-dictated notes → agent-enriched markdown → generated wikis → nightly background automation → agent-built graph visualizations. 21 minutes, AI Engineer conference talk. The pipeline: from scattered thoughts to browsable knowledge Goal: turn a jumbled Apple Notes folder into something navigable, interconnected, and discoverable — by you AND your agents Four stages: raw capture → enrichment → wikis → visualization Notes stay plain markdown; his demo app is Hubble (hub.md) — free, open-source, agent-accessible Stage 1 — raw capture: voice dictation wins Dictation is ~200 wpm — the fastest way to get thoughts down, no formatting, no polish Local options now: Handy (open-source, on-device model) or Voice Inc. ($20 lifetime; hotkey + mobile app — his pick) Scrappy, rambly notes are fine — volume of raw material is what LLMs need to work with later Stage 2 — enrich notes with an agent skill An “enrich note” skill: timestamp (so later passes skip already-enriched files), tags, source research, related-note search → backlinks Tags come from a controlled vocabulary in a reference file — instruct the agent to be reluctant to invent new ones (Claude “loves to get creative”) Runs in any harness — he demos Warp, notes open-weight models (GLM 5.2) work just as well; batch mode finds every un-enriched note and processes them all Stage 3 — generate wikis (the Karpathy idea) The concept comes from Andrej Karpathy’s “LLM knowledge base” gist: raw directory + focus area → generated wiki Example: an AI-news wiki grouping people, concepts, sources, and organizations (e.g. jazz musician Adam Neely on AI in music) Workspace use case: meeting notes → generated people/organizations section with backlinks to every related meeting Stage 4 — automate: run it on a cadence Enrichment and wiki generation take time — run them nightly/weekly in the background instead of on demand oz.dev (Warp’s cloud automation): a sandbox syncs your markdown via the Obsidian headless CLI (or git), runs the skill, syncs back Codex automations work too, but they need the laptop open — cloud runs while you sleep Stage 5 — visualizations: just ask an agent No tool to install — tell an agent to build an HTML/Tailwind graph view of your notes Clickable nodes, cluster view of what you actually care about, a “constellation” mode, habit/burndown charts Good for spotting gaps in your thinking, not just browsing “I wake up to a perfectly fresh wiki that I can review. It’s like the daily paper, but it’s your own.” ...