







As a hackathon project at our recent team offsite, Mati and I hacked together a quick experiment to test whether Bionic Reading actually works. Initial results suggest not, but there are some glimmers in the data justifying a revised experiment correcting the mistakes of our pilot study. If you have 10 minutes to read two Paul Graham essays, you can test Bionic Reading for yourself!
Does Bionic Reading actually work? We timed over 2,000 readers and the results might surprise you
Actually no, the results will probably not surprise you.

Decolonising Science Reading List
In the last couple of weeks, a lot of people have been asking me about what I’ve been reading as reference points for my commentary on the…

Guest post: If you’re going to critique science, be scientific about it
Loren K. Mell Editor’s note: This post responds to a Feb. 13 article in The Atlantic, “The Scientific Literature Can’t Save Us Now,” written by Retraction Watch cofounders Adam Marcus and Ivan Oran…

Are Scientific Papers Bad?
Better Science: A Reader
If you've been following Nintil -or been moderately curious1- you have probably come across claims that Science is broken in some way or another, that we live i…
Scientific Speed Reading: How to Read 300% Faster in 20 Minutes - The Blog of Author Tim Ferriss
How much more could you get done if you completed all of your required reading in 1/3 or 1/5 the time? Increasing reading speed is a process of controlling fine motor movement—period. This post is a condensed overview of principles I taught to undergraduates at Princeton University in 1998 at a seminar called the “PX […]
What happened to Science Goodreads and how do we rebuild it? A 65 million dollar question (at least) - Cosmik Labs
The story of the rise and fall of Mendeley
Science has been in a “replication crisis” for a decade. Have we learned anything?
Bad papers are still published. But some other things might be getting better.

Why We're Bootstrapping Readwise
Our mission here at Readwise is to improve the practice of reading through technology by an order of magnitude. While we have some strong opinions and good ideas (we think) on how to build such a product (indeed, we've already started!), what was less obvious was: How do we fund
Half of social-science studies fail replication test in years-long project
Results from massive, ‘eagerly awaited’ initiative reinforce concerns about the credibility of science — but raise hope for solutions.

Studying with Dwarkesh Patel - "Introduction to Quantum Mechanics" by Griffiths
Jason Shepherd on Twitter / X
I have quite a lot of thoughts on this..that won't fit in a tweet. Bottomline is that there are issues scientists in the US would love to improve - faster review, less grant writing, more stability and freedom to do risky science. The folks celebrating this doc are mostly../1 https://t.co/xlPAH9XYNC— Jason Shepherd (@JasonSynaptic) July 23, 2026
No. 003: Bixonimania hoax, Google's paper-writing agents, AlphaFold expands the frontier - The Hallway Track
The Hallway Track No. 003 - How AI is affecting the practice of science

How science found a way to help coma patients communicate
The long read: After suffering serious brain injuries, Scott Routley spent 12 years in a vegetative state. But his family were convinced that he was still aware – could a pioneering ‘mind-reading’ technique prove them right?

Scientific Agents: Evaluating Profession-Specific System Prompts on Scientific Tasks
Detailed profession-specific system prompts raise token use and estimated cost per response without a consistent accuracy gain. We evaluate Scientific Agents, an open-source corpus of 503 profession-specific AGENTS.md profiles, with Gemini 3.8 Flash via OpenRouter in the Pi agent harness. We compare matched profiles with four controls: a minimal baseline ("You are a helpful assistant"), the profile's opening role sentence, a generic scientific rigor guide, and a profile from an unrelated domain. Across nine text-based science benchmarks (4,531 sampled questions, 100 matched profiles), 4,488 items completed all five conditions after API-error retries, scored with automated, rule-based grading. The average profile-baseline accuracy difference is -0.6 percentage points (95% bootstrap interval [-1.5, +0.2] across fixed tasks), and no benchmark shows a statistically clear improvement. Matched profiles produced 1.5-2.3 times as many output tokens and cost 2.2-4.5 times more per successful call. On 60 tool-using BioMysteryBench bioinformatics problems (three runs each for baseline and profile), mean solve rates were 46.7% with the profile and 56.7% at baseline, a difference of -10.0 percentage points (95% interval [-16.7, -3.3]) driven by more frequent token- and time-limit stops under the profile. Longer prompts had one unexpected operational advantage: on SuperGPQA, frequent provider API drops left the short baseline with a correct first-pass answer on only 54.0% of items, against 71.6% with the profile. Generic and mismatched prompts were about as reliable, so this gain comes from prompt length or formatting rather than domain expertise. For the tested model and tasks, loading full profession profiles by default does not improve accuracy and costs considerably more; whether selective retrieval of profile sections or open-ended scientific tasks would change this remains to be tested.

The AI Chemist: To be trustworthy, LLMs need to show their work
Good scientists reveal how they do their experiments and report their results; so should any machine-driven research