iiinova — where your digital self keeps evolving

Dan Elton @delton.bsky.social · 04/10/2026

The white paper provides detailed protocols on how to administer modernized variolation at home safely. H2O2 is used to deactivate pathogens while preserving antigens, and antibiotic ointment is used as an adjuvant. white paper: radvac.org/modernized-v... Substack: substack.com/home/post/p-...

000

Dan Elton @delton.bsky.social · 04/10/2026

Here's a table showing the benefits of modernized variolation vs conventional vaccines, from the new white paper by Harvard-trained geneticist and @radvacproject.bsky.social co-founder Preston Estep and NIH NCI vaccinologist and virologist Chris Buck. (1/2)

110

Dan Elton @delton.bsky.social · 29/09/2026

OpenAI has been tracking how often their AI models are used for research. It's a hockey stick graph. We're living through a revolution in how scientific research is done, and it's happening suddenly and quickly. (from here: youtu.be/Fls_onRviPM?...)

000

Dan Elton @delton.bsky.social · 10/09/2026

BREAKING. Senator Josh Hawley just sent a letter to Sam Altman at OpenAI opening a probe into what transpired at OpenAI in the weeks leading up to the Hugging Face hack in light of new allegations. I think the public does deserve an independent investigation of what happened at OpenAI!

020

Dan Elton @delton.bsky.social · 09/09/2026

I found a ground truth dataset of western blots with known issues on Github, and will be testing our system on it. We aren't trying to compete w ImageTwin or ProoFig in terms of comprehensiveness (they do inter-paper etc), rather we're building a system which we plan to open source so anyone can run

131

Dan Elton @delton.bsky.social · 09/09/2026

Important bit I glossed over to fit in tweet: Part of the selection of the 1,000 was that one of the co-authors has a imaged-related retraction on a diff paper. For the next 1,000 we're doing the same sampling. Once we've improved the system more then maybe we'll run on a more random sample.

010

Dan Elton @delton.bsky.social · 09/09/2026

I tested the first iteration of my simple image duplication detection system on 1,000 biomedical papers. I made sure not to run on papers that are already retracted, have an image-related errata, or have image-related PubPeer comments. After review I found 5 hits total (0.5%). Will iterate now..

120

Dan Elton @delton.bsky.social · 09/09/2026

We didn't find any western blot issues, although I did see that some of these papers contained western blots. I'll test our system on blots with known issues to see if it can can actually pick them up... If you have any suggestions about where we should go with this, I'd love to chat!

110

Dan Elton @delton.bsky.social · 09/09/2026

I recently ran it on 1,000 papers. The papers have at least author that has a retraction for image duplication, but do not have a retraction, PubPeer comment, or errata. I am now reviewing.. so far I've made 4 PubPeer submissions. I've been tweeting out findings, will also go on Bluesky & website..

120

Dan Elton @delton.bsky.social · 09/09/2026

I built a system at @metascienceobs.bsky.social that does a thorough forensic audit of a scientific paper. Part of that is an image analysis module that checks for improper image duplication, using AI to help screen the "firings" from the detector (many duplications have innocent explanation) (1/2)

110

Dan Elton @delton.bsky.social · 03/09/2026

I have quibbles, but agree w a lot of this. Whether the international ban is practical remains to be seen. We can do a lot w narrow systems, like an "AI Einstein" that is really good at science but has no computer hacking ability and is not v good at understanding and manipulating human behavior.

000

Dan Elton @delton.bsky.social · 02/09/2026

i heard they may open source 1.2 soon, but I bet it will be a while before 1.3. what do you think?

030

Dan Elton @delton.bsky.social · 02/09/2026

A fast retraction!! It's a science miracle!! 😁

000

Dan Elton @delton.bsky.social · 02/09/2026

This is an actual book by Dan Ariely.. as advertised on his website back when it came out. 😅

120

Dan Elton @delton.bsky.social · 02/09/2026

Part 2 just dropped. Again I note that while there were warning signs in the paper, the smoking gun was in the raw data. How much fraud are we missing because raw data has been deleted or has been kept under lock and key? Should NSF call for scientists to release old data as a prereq for future $$?

010

Dan Elton @delton.bsky.social · 02/09/2026

There is always pushback from academics when fraud is publicisized. They worry about their own image being harmed, and as you know there is nothing academics care more about then their image / prestige / standing / status in the food chain... it's primal stuff from millions of years of evolution.

120

Dan Elton @delton.bsky.social · 02/09/2026

.. nearly 20 years ago to third party! One really wonders how much fraud there is that we are not able to catch because the raw data has been deleted or is kept under lock and key. At my nonprofit ( @metascienceobs.bsky.social ) we hope to find fraud with our AI agent systems, but we'll see...

020

Dan Elton @delton.bsky.social · 02/09/2026

yes, and a depressing thing from Data Colada's work is that you need the raw data files to find the smoking gun - not the SI/SM, the raw data. In this case there were clearly suspicious things in the paper, but the smoking gun was in the raw data, which they only got because Ariely had emailed it..

120

Dan Elton @delton.bsky.social · 02/09/2026

Claude Fable 5.1 just dropped today. Big jump in the benchmark for “agentic scientific research” 👀 Not so big jumps elsewhere, but some notably big decreases in cost! www.anthropic.com/claude-fable...

000

Dan Elton @delton.bsky.social · 02/09/2026

here's the replications we've been able to find so far at @metascienceobs.bsky.social. These come either from replication initiatives (we've imported several from psych) or from relatively simple keyword searches. I'll check out the paper you linked; we've been aware of the issue they highlight.

000

Dan Elton @delton.bsky.social · 01/09/2026

Always interesting to see how Data Colada finds fraud in psychology/behavioral science papers. The general theme is you need access to the raw data files to see the smoking gun. In this case, a third party, Kyle Hyndman, found the raw data in a 17 year old email Ariely had sent to him.

140

Reposted by Dan Elton

The Metascience Observatory @metascienceobs.bsky.social · 01/09/2026

Great news! You can now pip install our fetchPDF library! pip install fetchpdf it's the world's best command-line PDF-getter, using completely legal channels!

121

Reposted by Dan Elton

The Metascience Observatory @metascienceobs.bsky.social · 01/09/2026

Another Dan Ariely paper contains tampered (fraudulent) data, and it looks like Ariely was responsible (the file says "last saved by Ariely" and the only other author, Wertenbroch, claims he never had access to the dataset) datacolada.org/138

datacolada.org

[138] Artificial Deadlines (Part 1): Evidence of Fraud in an Influential Study About Procrastination

A new paper in Psychological Science (.htm) reports a failure to replicate Study 2 of Ariely and Wertenbroch’s influential article entitled, “Procrastination, Deadlines, and Performance: Self-Contr…

011

Dan Elton @delton.bsky.social · 30/08/2026

New on my Substack: how sleep optimization can backfire, five types of insomnia, and why I don’t recommend the Eight Sleep mattress cover. moreisdifferent.blog/p/how-sleep-...

moreisdifferent.blog

How sleep optimization can backfire

Plus, five types of insomnia and why I don’t recommend the Eight Sleep mattress cover

020

Dan Elton @delton.bsky.social · 29/08/2026

Incredible take-down. Go on Google and search “data center”. If you live near any population center you’ve probably been driving by data centers regularly without even noticing them.

130

Dan Elton @delton.bsky.social · 29/08/2026

That said, I do think some of the points made by AI skeptics & critics are still valid. For instance, LLMs can be tricked into giving the wrong answer if you give a common puzzle but change one thing slightly. And I'm skeptical they have robust generalizable world models with correct physics.

000

Dan Elton @delton.bsky.social · 29/08/2026

Although hallucinations are basically a non-issue now with most frontier models, they are unable to update on that because they are following anti-AI creators who still talk about hallucinations, and how AI is systems are trained to "mislead by making plausible text that's wrong".

100

Dan Elton @delton.bsky.social · 29/08/2026

Some of those folks who refuse to try modern AI (like Claude Opus 5, Claude Code) tried it a few years ago but ran into frustrating issues with hallucinations.

100

Dan Elton @delton.bsky.social · 29/08/2026

This might sound hyperbolic, but I think AI is the greatest tool for personal empowerment since the personal computer came out. Sadly, I’ve met people who refuse to try using AI because of all the fearmongering and negativity about AI that is rampant on both social media and in traditional media.

110

Dan Elton @delton.bsky.social · 29/08/2026

agree. But saying that "AI" (including agents like Claude Code) can't solve novel problems they haven't seen before or exhibit creativity has been proven false over and over.

010

Dan Elton @delton.bsky.social · 28/08/2026

(picture from a talk ~1.5 years ago by Joscha Bach when I hosted him in Cambridge, MA)

000

Dan Elton @delton.bsky.social · 28/08/2026

I feel bad for all the AI skeptics on this echo-chamber site who have their heads in the sand! I used to think "LLMs just interpolate except for rare cases of Grokking". I'm no longer so sure about that. And newer agentic systems are solving unsolved math problems, scoring very high on ARC-AGI, etc

330

Dan Elton @delton.bsky.social · 28/08/2026

With revenues skyrocketing there is no "AI bubble" overall at this time! Very different situation than the dot-com bubble period. There may be an over-extension of datacenter construction soon -- similar things have happened during every major infrastructure build-out previously. But not yet.

000

Dan Elton @delton.bsky.social · 27/08/2026

As a physicist, I have to admit I am guilty of occasionally saying that psychology is too descriptive. 😅

120

Dan Elton @delton.bsky.social · 27/08/2026

Again it's false it was just made up. If you have a specific issue with Heathers paper let me know. And Heathers said himself it was "highly unsystematic" in his paper.

000

Dan Elton @delton.bsky.social · 27/08/2026

I'd say my range is 5-15%. Somewhere in there. Really curious what you think.

000

Dan Elton @delton.bsky.social · 27/08/2026

Do you realize you're being a pompous jackass? You're not engaging in good faith with me or Heathers. I think 5-10% is probably about correct. Especially when factoring in that all of the lines of evidence Heathers presents are almost certainly under-estimates. Yes it is a hard reality to swallow!

100

Dan Elton @delton.bsky.social · 27/08/2026

What do you think is a better estimate? Note, I'm not doubling down. I acknowledge the original way it was stated was not good. I softened the statement to make it more accurate.

110

Dan Elton @delton.bsky.social · 27/08/2026

what are you referring to?

000

Dan Elton @delton.bsky.social · 27/08/2026

Also, saying that James Heathers just "made up" the 14% estimate is incorrect. If you read his paper, he has a very careful and logical line of reasoning to arrive at the 14% figure. Yes there are assumptions you can question, which Heathers acknowledges, but the estimate is not without merit.

120

Dan Elton @delton.bsky.social · 27/08/2026

I re-worded it. Also I'm calling the section "The data integrity problem" now, not "the fraud problem". It was on my mind to change both even before Carl Bergstrom criticized it. What do you think the true rate of papers containing at least one bit of fraud is in peer reviewed journals today?

100

Dan Elton @delton.bsky.social · 27/08/2026

How many humans are currently being paid to find and surface serious issues in existing papers? Like maybe six people worldwide? Studies show the current peer review system only catches ~30% of deliberately inserted serious mistakes.

000

Dan Elton @delton.bsky.social · 27/08/2026

It's a good point, and I was meaning to change that.. The "~" was doing a lot of work 😅. I agree the paper's titular conclusion conflates a lot of diff things (pre-vs post publication, anomalies vs proven fraud, etc). To his credit, Heathers acknowledged all that. Changed for now, may remove soon.

000

Dan Elton @delton.bsky.social · 27/08/2026

It seems worth a shot. What are you doing with your life other than making snarky uninformed comments on social media?

000

Dan Elton @delton.bsky.social · 27/08/2026

We are trying not to repeat mistakes of the past (most notably with YesNoError, which did publish a ton of slop loaded with false positives www.nature.com/articles/d41...). Note that was with o1, a dinosaur model by today's standards. Given recent advances in AI, we are optimistic.

nature.com

AI tools are spotting errors in research papers: inside a growing movement

Large language models are being used to check papers, but researchers warn they come with risks.

000