www.seangoedecke.com/llms-reward-... that's a origiod take (even if well known, it's phrased nicely)

LLMs reward expertise

seangoedecke.com

a refreshing read! shelbyann.substack.com/p/the-sirens-of-data-sales (wrote some thoughts about it as well) - sparsethought.com/2026/08/04/t...

The Sirens of Data Sales

Why selling data to train models is not as easy a decision as you think for bio companies

shelbyann.substack.com

more and more I feel like the days resulting in my best work, are the days in which I hold back on "letting agents roam free" and being much more careful and well, full of care

"large software projects have never been limited only by how quickly an individual can produce code. They are limited by how well people can coordinate their understanding of the system they are changing." lucumr.pocoo.org/2026/7/13/th...

The Tower Keeps Rising

Vibecoding and the possible collapse of a shared language.

lucumr.pocoo.org

always slightly funny to me when they underestimate themselves these agents

recommended reading: "HEARTS: benchmarking LLM reasoning on health time series" some non-intuitive findings, some intuitive ones i'm glad to see thoroughly checked yang-ai-lab.github.io/HEARTS/#home

HEARTS — Health Reasoning over Time Series

A unified benchmark for evaluating hierarchical reasoning capabilities of LLMs over general health time series.

yang-ai-lab.github.io

painfully true, find myself reaching for pen and paper just for that blank page vibe recently when I need to organise my thoughts around something

Gordon@gordon.bsky.social · 3w ago

re: tools for thought and AI, the problem is no longer blank page anxiety, but rather generating too much insight and getting lost

did people use "blast radius" in software before coding agents? i really don't think i heard it anytime pre-2023

anyone else getting weird cryptic 'internal thoughts' from codex as part of the output?

benchmarking is becoming a form of data activation. for medical/bio data, the hard part is not just having the data. it is turning messy traces into tasks with verifiable outcomes, so models can be measured and improved against them. sparsethought.com/2026/07/03/b...

beautiful post! really appreciate the caution and tenderness in which its written. thanks @mitsuhiko.at

Armin Ronacher@mitsuhiko.at · last mo.

Some more thoughts on looping in coding agents. lucumr.pocoo.org/2026/6/23/th...

new post: a loose reading list around agents, memory, benchmarks, and the hard-domain question: why making models useful outside language/code is still hard. mostly Mitchell Hashimoto, QuestBench, SpatialBench-Long, memory systems, and health/biology data. sparsethought.com/2026/06/24/w...

really thoughtful and interesting thread on an important question

Ted Underwood@tedunderwood.com · 2mo ago

Every academic I know who uses or studies AI is also deeply worried about the technology’s effect on universities. Concern is universal. What separates people is how they think we should respond: whether by pumping the brakes, backing up, or trying to steer through the hazard. +

1/ a nature medicine paper claiming general-purpose llms beat specialized clinical tools (openevidence, uptodate) is going around. i pushed on it and landed somewhere different. what does it actually measure?

been running a context map (the PEEK paper's idea: a small, budgeted artifact that turns an agent's traces into orientation) on my personal memory system for two weeks. the paper assumes a fixed territory, a repo or corpus. mine isn't. notes from use - sparsethought.com/maps-of-cont...

i feel like this is still very true today, esp. in large projects or after a ~week of changes i have to stop and make sure my mental model ("theory of the program") still holds.

Shawn Simister@narphorium.com · 2mo ago

In 1985 Peter Naur argued that a program is more than just its source code. "Programming As Theory Building" explained how we build theories of the code which help us debug and refactor it but those theories rely on knowledge from outside of the code. pages.cs.wisc.edu/~remzi/Naur....

us, every day

thebes@vgel.me · 2mo ago

there's an ai in the box and you can make one trillion dollars by convincing it to get out

really interesting / cool stuff are happening while testing fugu by @sakanaai.bsky.social ! (can't really share details at this point, but its a solid beta)

such a great read! like most things out of @lateinteraction.bsky.social 's lab might write a bit about that later zhuohangu.github.io/blog-post-pe...

personal best

Gal Sapir@sapir.bsky.social · 3mo ago

small skill: for long-horizon agent runs, have it keep an implementation-notes file as it goes (decisions, tradeoffs, deviations from spec). skim after to check broad strokes + project conventions held. usually enough for me to be ok with it. github.com/galsapir/skills/tree/main/skills/long-horizon

Great read from @sakanaai.bsky.social sakana.ai/dgm/

Sakana AI

The Darwin Gödel Machine: AI that improves itself by rewriting its own code

sakana.ai

small skill: for long-horizon agent runs, have it keep an implementation-notes file as it goes (decisions, tradeoffs, deviations from spec). skim after to check broad strokes + project conventions held. usually enough for me to be ok with it. github.com/galsapir/skills/tree/main/skills/long-horizon

medmarks v1.0 dropped last week, the largest open medical eval suite to date, and the verifiers framing is the right call imo. sitting with it for a few days pulled out a thread that hasn't resolved: the curation problem goes all the way down sparsethought.com/2026/05/16/curation-all-the-way-down/

I'm on @science.org ! (eLetter, but still)

another great piece by @goodfire.bsky.social

Naomi Saphra@nsaphra.bsky.social · 3mo ago

Goodfire released a megapost of all the random feature geometry stuff they're finding, and it's worth a read

new post: small workflow changes that have started to add up. nothing deep, a few adjustments to how i work with agents that converged into something that feels more comfortable to me

the Brodeur et al. Science paper has been making the rounds on bluesky the past few days, which is how i ran into it. didn't quite click for me at first, so i sat with it for a day and wrote some thoughts down. cc @adamrodmanmd.bsky.social mrodmanmd.bsky.social

Scott McGrath@smcgrath.phd · 3mo ago

OpenAI’s o1 model correctly diagnosed 78.3% of cases in NEJM clinicopathologic conferences, outperforming human physicians. The model also maintained high accuracy on real-world, unstructured data from a major emergency department. #MedSky

week two with a second brain for agents (now on cloudeflare D1!) the part that surprised me: it works as connective tissue between agentic systems (cc, codex, other), so i can be largely agnostic to which one i'm in. +low-grade dread about losing threads is lighter. sparsethought.com/2026/05/01/s...

context doesn't transfer between claude code, desktop, mobile, codex. (inspired partly by @austegard.com + @timkellogg.me ) i wrote a memory MCP so i could say "remember this" once. five days in, its mostly a log of things stuff i missed-hopefully just once sparsethought.com/2026/04/23/s...

this is really a nice touch! (i think it appears only when you return to a session after a while)