Serge Parel @spparel.bsky.social · 06/09/20263/ The model extracts and normalizes. A separate, deterministic rules engine makes the actual eligibility call. Full writeup, code, and adapter here: medium.com/@sparel/extr...medium.comExtraction is not enforcement: fine-tuning a local LLM for hERG assay curationThis is the first post in a series on applying QLoRA fine-tuning of local language models to pharmaceutical and agrochemical R&D workflows. 020
Serge Parel @spparel.bsky.social · 06/09/20262/ 7 out of 3,125 held-out records got wrongly admitted into a strict numerical QSAR set. A 0.224% false-admission rate. Fine for a chatbot, not for deciding what enters a training set. 100
Serge Parel @spparel.bsky.social · 06/09/20261/ Fine-tuned a small local model (Qwen2.5-7B, QLoRA) to curate hERG assay data from ChEMBL. The accuracy gain isn't the interesting part. The failure is. 110
Reposted by Serge ParelDerek Lowe @dereklowe.bsky.social · 31/08/2026The (alleged) dimethylmercury incident at MIT: thoughts based on whywe know so far:science.orgThe Alleged Dimethylmercury Incident 811332
Serge Parel @spparel.bsky.social · 31/08/2026Eastern cauliflower mushroom Chicken of the woods Giant polypore Tuberous polypore #MushroomMonday 030
Serge Parel @spparel.bsky.social · 22/07/2026No worries! Little Marco is on it: cybernews.com/news/europe-...cybernews.comMarco Rubio tells diplomats to play down talk of American tech “kill switch”US Secretary of State Marco Rubio has asked diplomats to push back against talk of a "kill switch" in American technology products following the White House's short-lived decision to keep foreigners f... 010
Serge Parel @spparel.bsky.social · 22/07/2026"What I don't like about [China's AI] is that it's all open source which means it's largely uncontrolled and not controlled in any way by us." - Eric Schmidt 110
Serge Parel @spparel.bsky.social · 19/07/2026Part 1, on the inference techniques behind it all: medium.com/p/your-gpu-i...medium.comYour GPU Is Mostly WaitingA field guide to LLM throughput in 2026: compression, speculation, and the art of never reading the same byte twice (Part 1 of 2) 000
Serge Parel @spparel.bsky.social · 19/07/20263/ Part 2 of my series covers the hardware ladder, the money math, and the boundary local models still cross badly: medium.com/p/the-year-l...medium.comThe Year Local AI Stopped Being a CompromiseModels, hardware, and the money math of running your own LLMs (Part 2 of 2) 100
Serge Parel @spparel.bsky.social · 19/07/20262/ The shift is physical: sparse MoE memory cost tracks active parameters, not total. A 35B model with 3B active runs 75-85 tok/s where a dense 27B manages 15-18. And against cheap frontier APIs, self-hosting breaks even around 3,000+ prompts/day, so you route by workload. 100
Serge Parel @spparel.bsky.social · 19/07/20261/ Three real machines from the past 90 days: a $1,740 garage rig on a decade-old server board serving a 30B coder at full speed, a MacBook running a 397B model streamed from SSD, and eight H200s running MIT-licensed weights that beat Claude Opus 4.8 on Terminal-Bench. 100
Serge Parel @spparel.bsky.social · 12/07/20263/ Plus the axis nobody benchmarks: agent workloads are bounded by prefill, not decode. Clearing accumulated context took first-token time from 31s to 2.5s. Full field guide: medium.com/p/your-gpu-i...medium.comYour GPU Is Mostly WaitingA field guide to LLM throughput in 2026: compression, speculation, and the art of never reading the same byte twice (Part 1 of 2) 000
Serge Parel @spparel.bsky.social · 12/07/20262/ They sort into three families: - move fewer bytes (rotation-based quantization) - read once emit many (speculative decoding, four generations in a year) - never pay twice (paged caches, byte-stable prompts) One study cut agent costs 87% with cache discipline alone. 110
Serge Parel @spparel.bsky.social · 12/07/20261/ Your GPU is mostly waiting. A 70B model at 4 bits is ~35 GB of weights, all read per token. At 128 GB/s of bandwidth that caps you at 3.5 tok/s. The teraflops never enter the equation. This is the memory wall, and every 2026 inference speedup is a way around it. 100
Serge Parel @spparel.bsky.social · 30/06/20263/ As models converge on quality, the next front isn't "who trained the smartest weights." It's who serves it fastest and cheapest. Users don't grade your inference paper. They're waiting for the answer. Full write-up: medium.com/p/dspark-the...medium.comDSpark: the DeepSeek inference upgrade that matters more than most model launchesMost AI headlines chase bigger models, smarter models, or more agentic models. DSpark is about none of those, which is exactly why it… 000
Serge Parel @spparel.bsky.social · 30/06/20262/ The bottleneck was never intelligence — it's the serving loop. DSpark uses a fast parallel drafter with a tiny sequential head bolted on (kills "suffix decay"), plus a scheduler that verifies more tokens when GPUs are idle and fewer when they're busy. 100
Serge Parel @spparel.bsky.social · 30/06/20261/ Everyone's chasing bigger models. DeepSeek shipped something quieter that might matter more: DSpark. No new flagship, no reasoning breakthrough. Just their existing V4 stack running 60–85% faster, with zero quality loss. How it works 🧵 100
Reposted by Serge ParelMrs. Betty Bowers @mrsbettybowers.bsky.social · 30/06/2026First, the White House doesn't date back to July 4, 1776, idiot. It opened on November 1, 1800. Second, the tackiest man in the United States and its president should *not* be the same person. Third, nice inspiration, fascist. 28054421466
Serge Parel @spparel.bsky.social · 28/06/2026“Die Schweiz gehe davon aus, dass die USA ihre Verpflichtungen aus dem Staatsvertrag einhalten würden” Unser Land wird von 🤡🤡🤡s regiert… 020
Serge Parel @spparel.bsky.social · 27/06/20268/8 Full breakdown, including the break-even math by usage tier and when it actually makes sense to buy vs wait: medium.com/@sparel/the-... #AI #semiconductorsmedium.comThe Mac Studio Quietly Lost 416GB of MemoryLast week Apple did something it almost never does. The company with the best supply chain in consumer electronics, the one that can… 010
Serge Parel @spparel.bsky.social · 27/06/20267/8 So: buy a local AI rig now, or wait? Honest answer: it depends entirely on what you spend on AI today. And the loudest "buy now" voices tend to be the ones holding memory stocks. 100
Serge Parel @spparel.bsky.social · 27/06/20266/8 Three firms make ~95% of the world's DRAM: Samsung, SK Hynix, Micron. All made the same call at once. No cartel needed when incentives align. A new fab takes 3+ years to build. Real relief doesn't land until ~2028. 110
Serge Parel @spparel.bsky.social · 27/06/20265/8 And it's locked in. Micron signed supply deals running to 2030 at floor prices "well above peak margins in any past cycle." Cloud giants prepaid billions to reserve memory before the wafers even exist. This is a contractual price floor, not a blip. 100
Serge Parel @spparel.bsky.social · 27/06/20264/8 The stat most coverage gets wrong: HBM takes ~22% of global wafer capacity but yields only ~9% of actual memory bits. Each HBM bit needs ~3x the silicon. So the real squeeze on consumer RAM is far worse than the "22%" headline implies. 100
Serge Parel @spparel.bsky.social · 27/06/20263/8 The cause is simple. AI data centers need a special memory called HBM. It's made on the same lines as the RAM in your laptop. So chipmakers choose: make HBM for Nvidia, or DDR5 for you? They pick Nvidia every time. The margins are 3-5x higher. 100
Serge Parel @spparel.bsky.social · 27/06/20262/8 DRAM prices rose as much as 98% in a single quarter. Apple raised Mac and iPad prices and admitted, in writing: "We have never seen a component price increase this much, this quickly." When Apple can't absorb it, nobody can. 100
Serge Parel @spparel.bsky.social · 27/06/20261/8 Earlier this year you could buy a Mac Studio with 512GB of unified memory. Today the top config maxes out at 96GB. The high-memory machines didn't just get pricier. They quietly disappeared. Here's what's happening with the 2026 memory crisis 🧵 101
Serge Parel @spparel.bsky.social · 14/06/20267/ Full piece, including the technical detail on the jailbreak, the precedent set, and why the next time it happens nobody outside the US gets a warning either: medium.com/@sparel/the-... #AI #Anthropic #AIGovernancemedium.comThe Night They Turned Off the Best AI in the WorldOn a Friday evening in June 2026, the US government picked up a phone and shut down the most capable AI model anyone had ever used. For… 010
Serge Parel @spparel.bsky.social · 14/06/20266/ What it means for everyone outside the US: any frontier model can now be pulled from global use within hours. Europe has no fallback. Most of the world has no fallback. Open-weight models just became infrastructure rather than experiment. 100
Serge Parel @spparel.bsky.social · 14/06/20265/ Earlier this year the same admin declared Anthropic a "supply chain risk" and tried to ban federal use of its products. Then kept using Anthropic models in classified Venezuela and Iran operations. "Vindictive" is the word this situation keeps reaching for. 100
Serge Parel @spparel.bsky.social · 14/06/20264/ Andrej Karpathy joined Anthropic in May to lead pre-training research. He's a US permanent resident, not a citizen. Under the directive, he can't access the model he was hired to improve. Same for co-founder Chris Olah. Same for Amanda Askell. The policy collapses on contact with reality. 100
Serge Parel @spparel.bsky.social · 14/06/20263/ WSJ, Reuters and The Information all confirm: Amazon CEO Andy Jassy called Treasury Secretary Scott Bessent and flagged Fable 5 as a security risk. Amazon is a $5B Anthropic investor. Read that sentence twice. 100
Serge Parel @spparel.bsky.social · 14/06/20262/ Claude Fable 5 launched on a Tuesday. By Friday it was gone. Official reason: a jailbreak that lets the model find vulnerabilities in code. Anthropic's response: that capability is already in GPT-5.5, which the government didn't restrict. So what actually happened? 110
Serge Parel @spparel.bsky.social · 14/06/20261/ 5:21 PM Eastern, Friday June 12. The US government picked up a phone and shut down the most capable AI model anyone had ever used. Globally. For everyone. Including the foreign-national researchers who built it. I spent the weekend digging in. 🧵 111
Reposted by Serge ParelNPR @npr.org · 01/06/2026A study finds that an mRNA vaccine is highly effective at preventing recurrence of this dangerous skin cancer, when used in combination with Keytruda, an immunotherapy drug. n.pr/4u3FHCZn.prA cancer vaccine made just for you. mRNA is back and it's fighting melanomaA study finds that an mRNA vaccine is highly effective at preventing recurrence of this dangerous skin cancer, when used in combination with Keytruda, an immunotherapy drug. 141234379
Serge Parel @spparel.bsky.social · 01/06/2026The caveat is the important bit: performance drops on raw, uncurated USPTO. Takeaway: reaction completion is not a side quest. It is part of the foundation for trustworthy reaction prediction, retrosynthesis, and process intelligence. 5/5 000
Serge Parel @spparel.bsky.social · 01/06/2026Their Constrained Reaction Balancer adds atom-balance constraints during decoding. It reaches 99.20% equivalence accuracy on random splits and 91.12% on extreme OOD splits. 4/5 100
Serge Parel @spparel.bsky.social · 01/06/2026The authors introduce CompleteRXN: ~200k aligned pairs of incomplete USPTO reactions and atom-balanced FlowER-derived targets, with random, mechanism-aware, and extreme OOD splits. 3/5 100
Serge Parel @spparel.bsky.social · 01/06/2026Core problem: open reaction datasets like USPTO are often chemically incomplete. The paper notes that only ~4.8% of USPTO reactions are atom-balanced. Missing byproducts, co-reactants, reagents, and stoichiometry are common. 2/5 100
Serge Parel @spparel.bsky.social · 01/06/2026Reaction prediction models are only as good as the reaction data they learn from. This paper is worth a look: CompleteRXN: Toward Completing Open Chemical Reaction Databases 1/5 arxiv.org/html/2605.00...arxiv.orgCompleteRXN: Toward Completing Open Chemical Reaction Databases 121
Serge Parel @spparel.bsky.social · 26/05/20265/ Token billing itself is transitional. Salesforce is already moving to "agentic work units" — paying for outcomes, not inputs. Three pricing eras: Flat-fee (2022–26): dead Per-token (2026–~27): now Per-outcome (2027–30?): coming Full breakdown ↓ medium.com/@sparel/ais-...medium.comAI’s All-You-Can-Eat Era Just Ended — And Most Companies Haven’t Heard the AlarmThe flat-fee AI pricing model collapsed in May 2026. A pricing war with hard July deadlines is reshaping the economics of coding agents… 000
Serge Parel @spparel.bsky.social · 26/05/20264/ The part nobody is talking about: 2 cost drivers are 100% within enterprise engineering control. ① Context bloat — sending the whole codebase on every call ② Routing every task to Opus when Haiku would do (5x cheaper) Estimated savings: ~$9.6M/yr for a 500-engineer org. 100
Serge Parel @spparel.bsky.social · 26/05/20263/ The mechanism is simple. Subscriptions were priced for humans (hundreds of prompts/day, bounded by sleep & attention). Agents have no such ceiling. One autonomous coding agent runs 100–1,000x more tokens per "seat." The plan was priced for the human. The agent showed up anyway. 210
Serge Parel @spparel.bsky.social · 26/05/20262/ May 14: Anthropic split Claude into "interactive" vs "agentic" pools. Same day: OpenAI offered 2 free months of Codex to switchers. April 27: GitHub announced all Copilot plans go usage-based June 1. Both pricing-war promotions expire in July. 100
Serge Parel @spparel.bsky.social · 26/05/20261/ Uber burned its entire 2026 AI budget by April. ServiceNow did the same. KPMG says U.S. orgs project $207M average AI spend over the next 12 months — nearly 2x last year. The all-you-can-eat era of AI coding tools is over. The deadline matters. 111
Serge Parel @spparel.bsky.social · 26/05/2026My take: SkillOpt is the safer near-term path for production agents. Before letting agents mutate core code, let them evolve compact procedural skills under strict eval gates. A skill without evals is just advice. Useful, but fragile. Paper: arxiv.org/abs/2605.23904arxiv.orgSkillOpt: Executive Strategy for Self-Evolving Agent SkillsAgent skills today are hand-crafted, generated one-shot, or evolved through loosely controlled self-revision, none of which behaves like a deep-learning optimizer for the skill, and none of which reli... 000
