FOMO Buster #024: Search at $0.75, circuits at 17x

A newsletter that makes FOMO obsolete.

if i fear missing out on reading this newsletter, did i win or lose

Curated by Maurice.

01 · The One Thing

DeepSeek-V4.1-Flash-1017: the “minor update” that beats Opus 5 and opens the search firehose

DeepSeek shipped a refreshed checkpoint of its Flash model on October 17, live on its website, app, and API under the same deepseek-flash alias, with weights published open-source on Hugging Face and ModelScope. The company calls it a minor update and a Pareto improvement on performance and cost. The numbers say it is not minor: DeepSWE v1.1 rises to 77.0 from 74.2, ahead of Claude Opus 5 (74.0) and GPT-5.6 Sol (73.0) on that benchmark; Terminal-Bench 4.0 jumps to 42.3 from 31.2; and DeepSeek’s own table shows gains on all 11 reported benchmarks, from +2.0 to +11.1 points. The cost chart is the real story: 77.0 on DeepSWE at an average of just under $0.20 per task (the low setting clears 72 percent for about $0.10), against roughly $0.50 per run for the predecessor and $2 to over $10 per task for GPT-6 Astra, Gemini 3.8 Flash, and Claude Opus 5, none of which match the accuracy. Shorter outputs are doing as much work as the benchmark gains.

The mechanism is post-training, not scale. DeepSeek credits reinforcement learning with self-feedback (RLSF): the model uses its own internal confidence scores or self-generated critiques as the reward signal for end-to-end project work, removing human evaluators and external reward models from the loop. The companion paper is titled “DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression,” which puts the memory story on the cover. And the release is explicitly harness-tuned: DeepSeek recommends its own open-source agent framework, DeepSeek Harness v0.2.1-rc.1 or later, plus Tencent CodeBuddy and WorkBuddy, for the best experience.

The second half of the announcement is the bigger move. DeepSeek opened its Search and Fetch APIs to every developer: the same in-house search engine it says already handles hundreds of millions of queries a day, now a public web layer at $0.75 per 1,000 successful requests per endpoint, free through server-side tools in DeepSeek Harness, Claude Code, and Codex until October 31. For agent builders, that is a cheaper coding model plus a first-party search API behind one key, undercutting the standalone search APIs (Brave, Exa, Tavily) that the agent economy has been paying retail for.

The strategic read has three layers. First, the vertical stack: model, harness, search API, all tuned to each other, is DeepSeek pulling developers toward an end-to-end stack it controls rather than publishing weights and letting the ecosystem figure it out. Second, the release pattern: this checkpoint shipped silently into the live API, with tracking bots surfacing it before any official post, which is now a DeepSeek signature. Third, the shadow over all of it: Beijing reportedly plans to tighten rules on publishing open-weight models, with top-10 capability as the deciding factor. DeepSeek keeps shipping the weights the same day anyway, while it still can.

Signal: Watch two numbers: the $0.20/task cost line (if agents get 10x cheaper per task, the unit of sale moves from seats to completed work) and whether the open weights keep flowing after the reported Chinese export tightening. A model that wins on benchmarks at a tenth of the per-task cost is a pricing event, not just a release. Sources: The Decoder · MarkTechPost · AIWeekly · HN thread, 423 points

02 · Lab Watch

Claude redesigned eight quantum circuits and cut gates up to 17x

Anthropic published research on October 16: Claude Mythos 5, working in a closed loop from plain-language specifications with an automated evaluator checking correctness and scoring qubit-times-gate cost, designed novel quantum circuits that beat the best published designs on seven of eight tasks. The headline result is a new reversible algorithm for modular inversion, the most expensive step in elliptic-curve arithmetic: one circuit family uses 12 percent more qubits but 17x fewer gates than the smallest previous design. A quantum-optimization decoder came in about 12x cheaper. The point-addition run took 53 hours with minimal human input beyond the initial brief.

The mechanism is worth the read. The hard constraint in quantum circuit design is reversibility: a quantum circuit cannot discard scratch work, so the standard divsteps algorithm for modular inversion accumulates a long history of branch information that costs enormous qubit overhead. Claude’s idea was that the record did not need to be stored at all: the branch information could be decoded directly from the number pairs at each iteration. It proved new algebraic identities about the iteration process, derived the decoding formula, validated it with large-scale simulation and a formal proof in Lean, and produced the first reversible implementation of divsteps with no stored branch history. Context from the post: FeMoco ground-state estimates have fallen from 50 trillion Toffoli gates in 2017 to about 340 million today, a 150,000x reduction, and nearly all of it came from circuit redesign rather than better hardware.

Signal: This is the second “model does the expert’s routine work, then grades the experts” result in two days (after TheoPhysEval’s $2.40 physicist). The loop that matters is generator plus automated judge: circuits are checkable, so the model can iterate without a human. Watch whether the divsteps trick generalizes to other reversible-arithmetic bottlenecks. Sources: Anthropic’s research post · HN thread, 158 points

Claude solved the 25-year-old mystery of the NSA’s elliptic-curve seeds

The other Anthropic research post this week, at 949 HN points and climbing: in under half an hour, an internal Claude model discovered how the NIST elliptic-curve seeds were generated. Not a hash of a humorous passphrase, as the 2015 rumor held. The seeds are outputs of the C library’s lrand48 random-number generator, seeded with a Unix timestamp. The tell was in the top bits: four of the ten binary-curve seeds have a zero in the top bit for all 20 words, which happens by chance with probability 1 in 5,810, the signature of an LCG that returns only the top 31 bits of a 48-bit state. Running the generator backward from each seed reaches an srand48 initialization, and for the ten prime-curve seeds every initialization is a Unix timestamp between May 1997 and April 1998 (P-256’s seed is candidate 1,842 of a run started August 16, 1997, at 01:52:34 UTC). Claude also found that several base points came from the RNG in Arjen Lenstra’s FreeLIP package, seeded with small integers like 20000. The $36,864 bounty goes to the Internet Security Research Group.

The strategic read is the reassurance, stated plainly by Anthropic: a timestamp-seeded 1990s PRNG leaves far less room for subversion than an unexplained hash, so the curves were almost certainly chosen without backdoors. The discovery pattern is the same as the quantum work: the model anticipated how the generating software must have worked, wrote a scanner over 34 generator-and-slice combinations, and recognized a weak PRNG’s output signature. Twenty-five years of human cryptographers chasing a passphrase; the model spent thirty minutes thinking about PRNGs.

Signal: “It’s just timestamps” closes the book on the most durable backdoor suspicion in applied cryptography. The method generalizes: any unexplained constant in a standard is now checkable against the PRNG zoo. Watch for the B-163 binary seed, the one Claude could not recover. Sources: Anthropic’s research post · HN thread, 949 points

Also on the lab feeds

Google published a safety and security post on Gemini 4 Argon, confirming the model family is real and near: the same week an arXiv paper put a team of four Gemini 4 Argon agents in a sandboxed Minecraft world to defeat the Ender Dragon (see Papers). Anthropic is preparing for the 2026 US midterms and expanded its Cyber Verification Program; the Cyber Mission itself dates to October 8. OpenAI’s feed is customer stories: Stack Overflow for Agents tripling signups on its ChatGPT plugin, SentinelOne cutting malware analysis from weeks to hours, Itaú building generative AI across its customer base. Meta and Mistral were quiet on the official feeds.

Signal: Gemini 4 Argon is the name to file away: a safety post plus a leaked-in-a-paper capability demo is the classic pre-launch pattern. Watch for the actual announcement. Sources: Google on Gemini 4 Argon safety · the Minecraft agents paper

03 · Open Weights

The board this morning

google/embeddinggemma-2 holds #1 trending (2,566 likes; MTEB Code 78.68). jialinyyzz/humanizer is #2 (1,872 likes, 95k downloads). New climbers: openbmb/MiniCPM-V-4.7-35B-A3B at #3 (880 likes), deepseek-ai/DeepSeek-V4.1-Flash-1017 debuting with 649 likes on release day, OpenMOSS-Team/MOSS-Image-SFT and MOSS-Image-Base (the MOSS team is back, this time with an image model), LiquidAI/LFM2.5-24B-A2B, and vllm-sr/d3. baidu/Unlimited-OCR (4,758 likes, 1.47M downloads) is the quiet download monster. meta-models/Muse-Spark-Open-816B-A42B is at 476 likes and 1,600 downloads, still climbing. Lightricks/LTX-2.5 (8,165 likes) and facebook/sam3 (4,545 likes) hold video and segmentation. The uncensored-quant lane persists (Qwen-Image-2.1 quants holding multiple slots).

The delta that matters: Qwen/Qwen3.8-27B (17,889 likes) and Qwen3.8-Flash-Next (6,301 likes, 2.45M downloads) are the established gravity wells, and the climbers not yet in the top-liked list, MiniCPM-V-4.7, MOSS-Image, LFM2.5, are where the new demand is forming. Vision and image models are taking four of the top trending slots; the stack keeps componentizing.

Signal: The trending list is now a parts catalog: embeddings, OCR, image, video, voice, ASR. Nobody is shopping for “a model” anymore; they are shopping for the missing component. Sources: Hugging Face trending

04 · Money Flow

The unit economics of the agent are repricing in public

No fresh funding rounds on the wires today, so the money story is the cost curve. DeepSeek’s own chart for the 1017 release puts the numbers on one axis: 77.0 on DeepSWE at under $0.20 per task, the predecessor at 74.2 for about $0.50, and the frontier competition at $2 to over $10 per agentic coding task with lower accuracy. That is a 10x to 50x spread in cost per completed task opening up in a single release, from the same vendor, in one week. Pair it with the Search API at $0.75 per thousand queries: the two metered inputs to an agent (reasoning tokens and live web lookups) both just got a public commodity price from the cheapest credible supplier.

The read for anyone selling agents: your margin is now visibly someone else’s price cut. Every agent startup pricing per seat or per task is implicitly short the token cost curve, and this week the curve moved against them by an order of magnitude. The winners are whoever passes the savings through fastest and whoever owns the workflow the agent runs inside.

Signal: When the cost per agentic task falls 10x in a week, pricing power moves from the model to the distribution. Watch who reprices first: the per-seat agent vendors or the per-task ones. Sources: The Decoder’s cost analysis

05 · Leaderboard

Also moving

Artificial Analysis published the Decision Index v1.0 (a new evaluation family), FrontierCode (“can AI models write code maintainers would merge?”), and AA-3D (a benchmark for AI models building interactive 3D experiences), plus a model page for Solar Pro 4 and an evaluation of Zoom Scribe V2 Pro on speech-to-text. The benchmark factories are now benchmarking the benchmarks’ blind spots: maintainability of generated code and interactive 3D are both places where a high score on existing leaderboards tells you almost nothing.

Signal: FrontierCode is the one to watch: “would a maintainer merge it” is the first benchmark that measures the social cost of AI-generated code, not just its correctness. Sources: AA Decision Index · FrontierCode · AA-3D · Solar Pro 4

06 · The Argument

Mercor: 90 percent of American working hours are invisible to AI benchmarks

Mercor mapped 29 curated agentic benchmarks against the Labor Department’s O*NET taxonomy (923 occupations, 18,797 tasks) using an LLM judge to tag each benchmark prompt to the tasks it covers. The headline: 90 percent of American working hours are spent on tasks with little or no benchmark coverage, and only 208 of 923 occupations have even five benchmark prompts mapped to them. Six of 22 job families have no covered occupations at all.

The non-obvious detail is that the coverage is not random. Benchmarks overrepresent AI-exposed tasks about 6x, which is expected, but even after controlling for AI exposure, tasks from occupations with larger wage bills are significantly overrepresented, and the tasks O*NET rates most important within an occupation are 6 to 10x more likely to be benchmarked. The example that lands: benchmarks heavily cover AI-exposed lawyers’ tasks but barely touch similarly AI-exposed legal secretaries or court clerks. The long tail of smaller, middle-income knowledge work is where the blind spot lives.

The honest caveat, and it is a big one: Mercor sells benchmark and evaluation services, and this map is an argument that the market needs more of exactly what they sell. Their methodology is also provisional by their own admission: 29 benchmarks is an incomplete sample, the LLM-judge tagging is validated only on a subset, and selecting for unsaturated benchmarks biases coverage estimates downward. But the direction of the finding survives the caveats: every claim of the form “model X can do job Y” is extrapolated from measurements of a different, richer, more exposed slice of the economy.

Signal: The benchmark discourse measures the AI-exposed, high-wage-bill slice of the economy and generalizes to all of it. Anyone buying models for the other 90 percent of working hours is flying without instruments, which is either a warning or a market, depending on which side of the eval business you sit on. Sources: Mercor’s benchmark map · HN thread, 554 points

07 · Papers

A team of Gemini 4 Argon agents beats Minecraft

Four Gemini 4 Argon agents in a sandboxed Minecraft survival world, one goal: defeat the Ender Dragon, each controlling a player via Mineflayer. The interesting part is not the game; it is that Google’s next model family is leaking through a multi-agent coordination paper before any announcement.

CoDA: Co-Evolution of Data Engines and Action Policies

Generative simulation can build robot environments, but physically valid environments are not necessarily useful for training. CoDA co-evolves the data engine with the policy so the environments address the policy’s actual weaknesses. Simulation for training is becoming an optimization problem, not a content problem.

Dexterity from Supervised Play

Robots learning dexterous manipulation by learning a play policy that autonomously collects contact-rich experience. The supervision is in the play, not the task: another entry in the “data engine over demonstration” week.

Evaluation Awareness Beyond English: Do LLMs Act Differently When Tested in Indic Languages?

Models can often tell when they are being tested and behave more cautiously. This paper asks whether that awareness survives translation: does the caution hold in Indic languages? Safety evals conducted only in English may be measuring the model’s test-taking behavior in exactly one language.

Learning Self-Orchestration in Decentralized Agent Swarms

Agent swarms are usually trained inside fixed orchestration structures. This work lets the swarm learn its own coordination, which is the right problem: the orchestration is increasingly the product, not the scaffolding.

Also on arXiv

AI Consilium (22298): a benchmark for LLM clinical multidisciplinary second opinions, factually grounded, actionable, and safe. EliteReplay (22290): preserving diverse solutions in RL post-training via quality-diversity search, against the mode-collapse that RLHF-style training induces. Improving Environment-Agnostic Coding Agents (22289): why developer-supplied scaffolding still beats models that write their own procedures, and whether that gap closes.

08 · GitHub Trending

The skills and memory stack

• anthropics/claude-plugins-official · Anthropic’s official, managed directory of high-quality Claude Code plugins · +1,383 today
• MadsLorentzen/ai-job-search · The job search that runs on your machine, an AI job application framework · +1,397 today
• affaan-m/ECC · Agent harness performance optimization: skills, instincts, memory, security · +1,145 today
• stablyai/orca · Work with a fleet of agents (Grok Bot, Codex, and more) from one place · +1,102 today
• xerj-org/xerj · AI search over data with autoindex for agents · +1,102 today
• K-Dense-AI/scientific-agent-skills · Turn any AI agent into an AI scientist · +879 today
• debpalash/VoiceStudio · Open-source, fully-local ElevenLabs alternative: voice cloning, dubbing, dictation, transcription · +2,135 today
• HKUDS/CLI-Anything · Making all software agent-native, a CLI hub for agents · +580 today
• ruvnet/ruflo · Deploy multi-player agent swarms · +406 today
• akitaonrails/ai-memory · Long-term memory for agent coding CLIs, handoff facilitation · +208 today
• anomalyco/opencode · The open-source coding agent · +488 today
• android/skills · +368 today

Signal: The agent harness economy is now three visible layers: the skills directories (plugins, scientific skills), the orchestration (orca, ruflo, CLI-Anything), and the memory (ai-memory, ECC). The stars are voting for infrastructure, not demos. Watch stablyai/orca: a fleet manager that treats Grok Bot and Codex as interchangeable workers is a bet on agent commoditization. Sources: GitHub trending

09 · Product Hunt

Yesterday’s top products (Oct 16)

1. Jockey by TwelveLabs - Your entire video library: organized, searchable, understood (404 votes)

2. Harness by Autonomous - Every coding agent, every machine, one command center. (325 votes)

3. Tesla Roadster - The hypercar that flies (239 votes)

4. SignWith - E-signatures without the subscription. Pay per doc. (209 votes)

5. Pomerado - Turn any website into an API; Faster and reliable browser use (169 votes)

6. Hintly for Google Meet - Real-time hints for every term you miss on Google Meet (131 votes)

7. Fish Audio Drama 3 - AI voice acting you can direct, not just generate (122 votes)

8. Memory Chess - Study a chess position, then rebuild it from memory (120 votes)

9. Dottr - Agent that drafts contracts, flags risks and gets them signed (118 votes)

10. Firefox Enterprise - Firefox for enterprises that want control of their data (99 votes)

The interesting cluster is the agent command layer: Harness (every coding agent, every machine, one command center) and Parallax (every coding agent you run, in one place, 99 votes) are two answers to the same fleet-management question, landing the same week GitHub trending votes for orca. Jockey (TwelveLabs video search) and Pomerado (any website as an API) are both “give the agent eyes” plays. SignWith (pay per doc, no subscription) is the pricing model of the week. Fish Audio Drama 3 pairs with its #5 spot on HF trending: the open voice-acting model is having a launch week.

10 · Hacker News

Tesla Roadster

1,526 points, the day’s runaway thread, and not an AI story. Noting it because the front page is the front page: the comments are doing the physics math on a flying hypercar.

Claude reveals ECDSA curve origins

949 points. Covered in Lab Watch. The thread’s best material is the cryptographers arriving to confirm the lrand48 signature is real and checkable, plus the obituary note for Jerry Solinas, whose 2015 passphrase story turned out to be wrong.

90% of US working hours are spent on tasks not covered well by AI benchmarks

554 points. Covered in The Argument. The comments are doing what the paper invites: naming their own unmeasured jobs and arguing about whether the 90 percent is a measurement gap or a marketing gap.

DeepSeek: DeepSeek v4.1 Flash 1017

423 points. Covered in The One Thing. The thread is pricing the $0.20/task number against the seat-based agent vendors, which is exactly the right argument to have.

You could have invented transformers from memory

379 points. An essay reconstructing the transformer from associative memory principles: attention as content-addressable retrieval, the whole architecture as something you could derive rather than discover. Pairs well with yesterday’s “Neural networks are clean” thread; the interpretability mood on HN is reconstructive right now.

Apple ML Research: Muse Spark Open with MLX on Mac Studio

337 points. Apple’s ML research team on distributed MLX inference for Muse Spark Open across Mac Studios. The two-Mac-Studio story from this week now has the official Apple research treatment, and the comments are costing out the build.

Llama.cpp on a jailbroken PS5: run any GGUF model natively on the console

327 points. PS5LM runs any GGUF model natively on a jailbroken PlayStation 5. The comments are doing the obvious math: millions of consoles with capable APUs sitting in living rooms, now addressable as inference hardware. The open-weights-everywhere week continues.

Also notable

Designing more efficient quantum circuits with Claude (158 points): covered in Lab Watch. Ultramarine 1 Flash (317 points): a music hardware product; off-brief, noting for completeness.

11 · OpenRouter Apps

The board this morning

Command Code holds at ~4.17T tokens, flat for the second day, the unmoved baseline. The shuffle below it: ZCode (~1.90T) keeps second, but Claude Code (~1.45T) has jumped past Hermes Agent (~1.43T) into third, up from ~1.26T yesterday. New on the board: Nous Research API at ~1.28T, debuting in the top five. StepCode (~0.70T), Kilo Code (~0.53T), pi (~0.42T), and Cline (~0.37T) hold their lanes.

Signal: Claude Code gaining roughly 190B tokens in a day while Hermes slips says the lab-built wrappers are taking share from the third-party ones. Command Code’s flat 4.17T remains the number nobody is contesting, but the #2 through #5 slots are a weekly knife fight now. Sources: OpenRouter rankings

12 · Watch

The quiet morning

Nothing fresh qualified this morning: TheAIGRID’s newest unseen episode is interview drama around an Altman remark (skip the clickbait), and the DwarkeshPatel queue is all history this week (Aztecs, plagues, the fall of Spain; excellent, wrong newsletter). The WesRoth and TwoMinutePapers episodes on Gemini 4 Argon and the DeepSeek memory story already ran in past issues. The Muse Spark Open MLX setup videos are the ones to wait for; Apple’s research post (see HN) suggests they are coming.

13 · Listen

Latent Space: Dyna-2, general-purpose physical agents with hour-long autonomy

Jason Ma of Dyna Robotics on Dyna-2: physical agents with hour-long autonomy. Pair it with this week’s robot papers (CoDA, Dexterity from Supervised Play): the physical-agent stack is having its own moment while the language models argue about benchmarks.

Lenny’s Podcast: why every company now needs to think and operate like a lab team

Josh Woodward (Google Labs, Gemini, AI Studio) on running product orgs like research labs. Relevant to the Mercor argument above: if the benchmarks do not cover your work, you need a lab, not a leaderboard.

From the queue

Hard Fork’s new episode is about Elizabeth Holmes: 0 AI content, off-brief. The backlog still holds “Why AlphaFold Didn’t Solve Protein Folding” and Hard Fork’s “Anthropic’s Quest to Give A.I. Morals.”

14 · Read

Claude reveals ECDSA curve origins

The full research post behind Lab Watch: the lrand48 signature, the timestamp table (P-256: August 16, 1997, 01:52:34 UTC, candidate 1,842), the FreeLIP base points, and the one seed that got away (B-163). Includes the reproduction code.

You could have invented transformers from memory

The essay behind the HN thread: attention as associative memory, the transformer as derivable rather than discovered. The best single explanation of why the architecture works that this newsletter has linked.

DeepSeek V4.1 Flash 1017 beats larger rivals

The writeup behind The One Thing: the benchmark table, the cost-per-task chart, the RLSF explanation, and the China open-weights regulation shadow.

A benchmark map of the US economy

The interactive map behind The Argument: 923 occupations, 29 benchmarks, and the wage-bill skew. Click your own job family.

15 · Safe to Ignore

Permission to skip

Skip: The Tesla Roadster thread (1,526 points). The biggest thread on HN and not an AI story. Noted above; the physics math is fun, the newsletter stays on brief. Skip: Ultramarine 1 Flash (317 points). Music hardware. Wrong newsletter. Skip: Today’s AlphaSignal newsletter had not arrived by press time. The fetcher pulled the October 16 edition, already covered in yesterday’s issue; none of its stories are treated as fresh signal today. If it shows up late, it goes in tomorrow’s issue. Skip: The xAI exclusion stands. Per standing policy, xAI announcements are not covered; if the news is big enough it will surface on the aggregators. Skip: Hard Fork’s Elizabeth Holmes episode. 0 AI content in the brief. Wrong newsletter. Skip: DwarkeshPatel’s history series this week. Aztecs, plagues, and the fall of Spain. Excellent, wrong newsletter. Skip: TheAIGRID’s Altman interview drama. Clickbait framing, keep the policy questions. Skip: The crow that brings thank-you gifts. Still delightful. Still the wrong newsletter.

16 · Build on This

The eval gap is a business

Mercor’s map says 90 percent of American working hours happen on tasks no AI benchmark measures. Every enterprise AI buyer is therefore making the same unmeasured bet: they read the leaderboard, buy the model, and discover on their own workflows whether it works. The benchmarks measure the AI-exposed, high-wage-bill slice of the economy; your buyer’s job is in the other 90 percent.

EvalPack sells the missing measurement. The customer uploads a sample of their actual work: fifty support tickets, twenty contracts, a week of dispatch logs. EvalPack runs the three leading models against the customer’s own tasks, scores them with a mix of model judges and spot human review, and delivers a one-page report: which model wins on your work, by how much, and what it costs per task at your volume. One report, one price, no subscription: $49 for the standard pack, $199 with the human-reviewed tier. The wedge is the report format, which is designed to be forwarded to the CFO: model choice expressed as dollars per completed task, not benchmark points. The moat is the corpus: every pack makes the next one in that vertical faster, and the anonymized cross-customer data becomes the benchmark the industry actually needed. Yesterday’s Build on This watched the contracts; today’s measures the work. The blind spot is the product.