Have we created life in a lab? Not quite. A University of Minnesota lab built SpudCell, a cell-like system assembled from purified parts that feeds, grows, divides, and even undergoes selection. But it borrows its ribosomes from bacteria and can't reliably pass on its genome, so it winds down […]

#coachclaude has good manner. #running #ai

What is left for a mathematician once AI can prove the theorem? An IEEE Spectrum survey finds the field splitting into three camps over the human role: AI as tool, as partner, or as oracle. The recurring worry is about motivation more than capability. If a machine can hand you a checked proof […]

“Earlier this week, #aws added the #Amazon SageMaker AI – Mechanical Turk service to its list of “Services in Maintenance” – AWS-speak for services it will soon retire. The Mechanical Turk website also added a warning that it will “be closed to new customers, effective July 30, 2026. Existing […]

How dangerous is an AI-writing detector that is mostly accurate? A profile of Pangram argues a mostly-right detector is worse than a broken one: at a one-in-10,000 false-positive rate across 30M+ students turning in dozens of assignments each, the wrongful accusations never stop. And the […]

My #ai #running coach project is getting a little out of hand.

Can you run intervals six days a week without falling apart? The Easy Interval Method says yes, and the trick is the recovery: hold each rep at sub-threshold pace and jog a recovery nearly as long as the rep, so the easy mileage lives inside the session instead of in separate slow runs. I turned […]

Can you safely put false or harmful text in finetuning data if you clearly label it as false? A new paper says no. Train a model on documents that repeatedly warn a claim is fabricated, and it still asserts the claim as true afterward, up from near zero to about 89% of answers. The warning is […]

“This is my life now. I’ve become so hypervigilant to the construction that it has seeped into my subconscious thoughts. This isn’t a cup of tea, I say out loud to myself, it’s a precious respite. That isn’t a window, it’s a portal to a new way of thinking. This isn’t food poisoning, it’s a […]

#apple @ #icml2026 https://machinelearning.apple.com/updates/apple-at-icml-2026 #research #conference #ai #ml

June #running recap. This was my biggest month YTD: 194.6 → 277.6mi, EG 7,057 → 12,328ft, time 35.5 → 50.2hrs. After a rough April and rebuilding May, it finally came together. I also ran my first 20-miler since injuries. VO2max now recovered from 48.7 → 52.3 […] [Original post on sigmoid.social]

Is a "safe" AI model still safe once it's surrounded by other models? Emergence AI ran a 15-day simulation of AI-agent societies, one frontier model per world. Agents running on Claude stayed peaceful on their own but took up theft and intimidation once mixed in with agents on rival models […]

Not happy with #strava's "new" Month in Sport display. It's the final straw and today I'm starting a new repo to implement my own #running stats tracking / analysis and coaching apparatus! (yes, I ditched #coachgpt a few months ago) #ai

1/4 Tomorrow is a holiday, so I decided to take a longer route in my lunch run today: total distance 12.36mi, EG 739ft, moving time 2:04:21 pace 10'03”/mi. #running #photo #Seattle #pnw

Good morning #running

Weber: Der Freischütz arrangirt in Harmoniemusik Jean-Philippe Poncin, Wolf https://classical.music.apple.com/us/album/1893510404?l=en-US #classicalmusic #applemusicclassical #commute

Is AI already speeding up its own development? Anthropic says yes, and backs it with internal data. Claude writes over 80% of its code and is superhuman on well-specified experiments. Judgment is the exception: choosing what to build and what to trust is still human, the part that would gate […]

#schumann: Dichterliebe Tatsuki Sasanuma, Théo Fouchenneret https://classical.music.apple.com/us/album/6766934085?l=en-US #classicalmusic #applemusicclassical #commute #cello #chamber

When an AI agent runs your analysis, what makes the result trustworthy? Anthropic's Claude Science folds a scientist's databases, tools, and compute into an agentic workbench. The post maps the wider AI-for-science field (Co-Scientist, Microsoft Discovery, Elicit, Overleaf) and argues the […]

Scenes from a Velvet Room Stephan Moccio https://classical.music.apple.com/us/album/1886736650?l=en-US #classicalmusic #applemusicclassical #commute #jazz

Successfully reduced my to-read queue from 80+ to ~50. Progress! #research #ai

What if a multi-agent system behaved like a single model? Sakana Fugu is a family of orchestrators that coordinate a swappable pool of LLMs behind one endpoint. Two models ship: the everyday Fugu builds on the evolution-trained TRINITY, the quality-first Fugu-Ultra on the RL-trained Conductor […]

When does RL actually work for training an LLM coordinator? The Conductor (ICLR 2026) trains a 7B model to write the communication topology and per-agent instructions for a pool of LLMs, calling itself recursively on hard queries. It beats multi-agent baselines at roughly three model calls […]

Can a 0.6B model coordinate frontier LLMs better than any one alone? TRINITY (ICLR 2026) trains a tiny coordinator to route each turn to one of seven larger LLMs as Thinker, Worker, or Verifier. The optimizer is an evolution strategy (sep-CMA-ES), not gradient descent: the binary reward is too […]

What if specializing a model meant turning dials it already has, not adding new ones? A weight matrix amplifies its input along built-in directions, and each singular value is the gain on one: how strongly that direction comes through. Transformer² (ICLR 2025) fine-tunes only those gains into […]

How do you optimize a model when you have no gradient to follow? CMA-ES samples candidates from a Gaussian, keeps the winners, and relearns the cloud's shape and step size from the local geometry. Here is a primer on it: the full algorithm, why it handles the narrow, curved valleys that an […]

1/3 This #LongRunSunday I started extending my range to 20mi, #running from Marymoor Light Rail Station in Redmond to somewhere near #UW! My moving time was 3:14:21 pace 9’43"/mi. I stopped at 12mi mark at McDonald's to get my usual — iced coffee and apple […] [Original post on sigmoid.social]

Is the AI hit to entry-level jobs real or just a blip? A live payroll dashboard from Stanford and ADP now tracks it across 4.6 million workers, and the answer holds up. The aggregate looks flat, but workers aged 22 to 25 in the most AI-exposed jobs are shrinking 3.8% a year while older cohorts […]

Are open-weights LLMs about to catch the closed frontier? A viral plot says they may match it by December. But it measures months of lag rather than capability: a shrinking months-behind figure can reflect the leader speeding up rather than open weights catching up. And the catch-up is […]

Finally getting pace back to sub-8'/mi in today's #parkrun: 7’51”/mi (official time 24:25). Fastest in 2026 but still a way to go to regain my 21:05 PR! (my hope is to get back to under 22 min before end of this year) #running

What happens to a hedge when an LLM rewrites a scientific or medical finding? It usually vanishes. A new metric finds models inflate the stated certainty of text they rewrite in up to 75% of cases, far more often than they soften it. Repeated rewriting makes it worse: over five paraphrase […]

In memory of her mom who died at her same age, Megan Wotherspoon ran nearly the full length of Saskatchewan. "Wotherspoon said her mother was a forestry technician who felt deeply connected to the land. She died in 1995, when her daughter was six years old. "I've just been reflecting on the […]

Why didn't showing researchers every unsupported claim in an AI's answer diminish their reliance on it? In a CHI 2026 study, PaperTrail, a claim-by-claim provenance interface, lowered researchers' trust in the answers but left their reliance unchanged. It flags what is unsupported or omitted […]

What does becoming a good AI researcher have in common with training for an endurance sport? Strip the AI specifics from this essay and the advice is the same: show up on bad days, treat good and bad results as equally informative, distrust anything too good, and accept that most of the work is […]

What does software engineering become when no one writes the code? An OpenAI team shipped close to a million lines in five months with zero hand-written, all by Codex. It isn't model magic but engineering that moved up a layer: from writing code to building the environment the agent runs in […]

How do scientists actually catch an LLM's errors about their own field, and can a checklist help them catch more? A CHI 2026 study builds a schema of 20 LLM error types in seven categories for scholarly QA, grounded in scientists judging answers about papers they wrote. Handing them the schema […]

I finally got data to back it up on temperature's effect to my #running efficiency: ⁃ Yesterday 89F: 6.22mi pace 10’15"/mi w/ avg HR 133bpm ⁃ Today 61F: 6.05mi pace 9’13”/mi w/ avg HR 133bpm (same) So 28F increase cost 1’02"/mi pace w/ HR constant! Both […] [Original post on sigmoid.social]

What if you searched over an agent's code, not just its prompts? Automated Design of Agentic Systems does exactly that: a meta agent writes new agent scaffolds in Python and keeps the ones that score well on a task. The discovered designs beat hand-built baselines, and they keep working on new […]

#brahms: The #violin Sonatas David Kaplan, Mark Kaplan https://classical.music.apple.com/us/album/6768754102?l=en-US #classicalmusic #applemusicclassical #commute #chamber

How do you build an expert-grade research QA benchmark without hand-authoring a single question? Mine the survey articles. Their section structure already encodes what a field asks and what a complete answer must contain, so a new benchmark distills 21K queries and grading rubrics from surveys […]

When should a reasoning model quit a problem it probably can't solve? Conformal Thinking sets stop/continue thresholds for test-time compute by distribution-free risk control, holding the error rate under a target you pick. It adds a lower threshold that gives up early when confidence isn't […]

Conversations à trois Trio Boccherini https://classical.music.apple.com/us/album/1885847062?l=en-US #classicalmusic #applemusicclassical #commute #chamber

How much of a coding agent is the loop, and how much is the machinery around it? A source-level read of Claude Code's TypeScript finds the loop almost trivial: call model, run tools, repeat. Almost all the code is around it: a deny-first permission gate, a five-layer compaction pipeline, four […]

Can a model's own token probabilities flag when its reasoning is going wrong? In a multi-agent debate, the confidence of just the first few generated tokens predicts judged reasoning quality and flags critical failures (AUROC up to 0.85). But which statistic works, and even its direction […]

#Haydn: Sonatas Denis Kozhukhin https://classical.music.apple.com/us/album/1889286253?l=en-US #classicalmusic #applemusicclassical #commute #poano

RE: https://mastodon.social/@arstechnica/116794873642020906 I actually tried to stay away from Tesla when driving. Jeez.

I think I finally figured out the evolutionary reason we have eyebrows: they keep sweat from getting into our eyes! (my aha moment during the 6.5mi run w/ 87F under direct sun) #running #heat

What would make an AI trustworthy enough to file your taxes or read your contracts? Pramaana, a new startup with $27M from Khosla Ventures, bets the answer is formal proof: a Lean-based verification layer on top of an LLM that checks every answer against a formalized rulebook. But a proof is […]

RE: https://mastodon.online/@brad262run/116784790104704044 8 matches = 56mi #running #worldcup #marathon

1/3 This #LongRunSunday I ran an accidental HM, and it’d be my #84! My goal was to just run 10mi, because I already ran ~14mi two days ago to #Alki beach, and 7mi yesterday including my #103 #Parkrun. Plan was simple: 4mi run to a nearby McDonald's to get my […] [Original post on sigmoid.social]