New York / Lisbon
Joined July 2009
- Tweets3.2K
- Following1.1K
- Followers606
- Likes12.1K
Andreas Borg retweeted
We are bringing agents to the wet-lab bench at Google Research to navigate complex biological landscapes. Learn how we are grounding agents in physical lab robotics to run fast closed-loop molecular design cycles, ultimately discovering single-digit nanomolar protein binders for cancer targets →goo.gle/4B5mMMB
Andreas Borg retweeted
We are giving away 1 MILLION LigandForge designs as part of the Adaptyv Proteinbase + Anthropic protein design competition.
challenge.ligandai.com
Andreas Borg retweeted
No, "anthropomorphizing" is a technical problem, about how AI works, not something like woke language policing. We say it's wrong because it's "technically" wrong not "morally" wrong.
If you fail to set the brakes when parking on a hill, and your car then rolls down the hill, you don't claim that the car "decided to drive away on its own". Such anthropomorphizing distorts what really happened -- that you failed to set the brakes.
The same applies to AI. When you "anthropomorphize" the AI you fail at understanding what actually happened.
Andreas Borg retweeted
Everyone's building a "second brain" now.
Obsidian, Notion, 400 notes, a connection graph that looks like a subway map.
And you still can't remember what you read yesterday, because you spent the whole day scrolling instead of thinking for even 10 minutes.
Why do you need a second brain when you barely use the first one?
Andreas Borg retweeted
I'm going to start a licensing program for bio dooming where we march applicants into a lab and force them to stitch together a 10kb construct from oligos with only a chatbot as a coach.
To make it realistic we'll employ jaded postdocs as proctors to belittle them ruthlessly.
Andreas Borg retweeted
BREAKING: OpenAI might have stolen another major proof.
In a detailed Mastodon post, which I report in full in the comments, Andreas Thom presents several pieces of evidence suggesting that OpenAI may have trained Astra on conversations in which he and Gábor Kun were working on Gromov’s soficity conjecture, one of the ten problems OpenAI later announced Astra had solved.
I know Andreas. We met several times early in our careers. He is an exceptional mathematician, a leading expert on sofic and hyperlinear groups, and one of the most respected scholars in the field. He has spent two decades working on this problem.
If his account is correct, this is not a minor dispute over attribution. It would mean that unpublished human work was absorbed into a model and then presented to the world as a breakthrough by the model itself.
And if the allegations raised by Levent Alpöge, Tristan Buckmaster, and now Andreas Thom are all substantiated, we are no longer looking at isolated incidents.
We may be looking at one of the greatest intellectual scandals in the history of science.
AI is not discovering new mathematics.
AI is stealing human discovery.
Andreas Borg retweeted
a nice demonstration of Claude Science, but worth clarifying that the design is not "done by Claude" but by orchestrating tool calls of open-source, task-specific protein design models: PXDesign, RFdiffusion, Genie, BoltzGen, etc
I think the direction of LLMs using biology-specific models is a good one!
Many drugs work by binding to a specific target in the body and blocking or changing what it does. An important first step in the drug development process is designing a molecule that can bind tightly to its target. Traditionally, that's meant weeks or months of expert work per target, sifting through a large number of candidates to identify the few that work.
We wanted to test if Claude could successfully design novel protein binders from scratch (also called de novo design). With a protein design prompt written by a human expert, Claude autonomously designed protein binders against 14 out of 15 targets.
We then worked with Adaptyv Bio and Twist Bioscience, who independently built and tested the proteins Claude designed.
Andreas Borg retweeted
Prediction: Bio is about to become an even more popular college major. Domains where answers can be formally verified are going to fall like ninepins.
In the case of bio, frontier models do ok at protein structure, molecular interactions and some genomic predictions but still struggle with context dependent behavior at the cell and organism level. That will be the next ultimate challenge.
An internal version of Astra, @OpenAI’s next major model family, solved 10 major open problems in mathematics, quantum complexity, and theoretical computer science.
We believe it will be a major step for scientific reasoning. openai.com/index/ten-advance…
Andreas Borg retweeted
The Genetic Agency Technology Conference (GATC) was born out of the idea that genetic agency – an individual’s ability to take action at the genetic level to live a healthier life – is a collective responsibility, shared across the gene therapy ecosystem.
@PiteriSamantha reflected on the unique structure of last year’s meeting, which was intentionally designed to foster small-group discussions, align goals across organizations, and surface unexpected partnership opportunities that could help move the field forward.
Andreas Borg retweeted
How much better could AI agents get at scientific discovery if they could search in biological embedding space instead of text?
Over the past few months, we’ve been building Ladon, our bulk transcriptomics foundation model. As an experiment, we embedded over 1 million public RNA-seq samples spanning 20+ years of research, exposed that embedding space through an MCP, and let Claude search it.
Today, AI agents search titles, abstracts, metadata, and keywords. They can’t search the biology itself.
Ladon maps every RNA-seq sample into a shared biological embedding space, allowing agents to search by biological similarity instead of text. A naive search mostly retrieves studies from similar tissues or diseases. What we actually want to match is the transcriptomic signature itself: the pattern of what changed in a study, not the tissue or disease it happened to use. Searching on that lets an agent connect experiments across diseases, tissues, and perturbations, even when the papers share no keywords.
We explored a case study where this surfaced novel hypotheses that would have been difficult to find through traditional search with agents.
Read more in the blog: blank.bio/ai-agents-will-acc…
If you're sitting on large RNA-seq datasets and want to harmonize them with the public transcriptome to power agent-driven discovery, we'd love to collaborate. Reach out!
Andreas Borg retweeted
Heard about this coming at last year's GATC! I'm excited for these milestones that bring us closer to genetic agency: scaling up treatment of genetic therapies and offering hope to patients with few and often no good treatment options.
Experts in gene editing, n=1 medicine development, in vivo delivery, rare diseases, and cell & animal models of genetic disorders have founded the Center for Therapeutic Genetics (CTG), a non-profit effort involving @broadinstitute @BostonChildrens @jacksonlab, as well as others from industry, academia, and patient advocacy organizations, to bring the promise of programmable medicines including base and prime editors to patients suffering from rare genetic diseases. This article from @ginakolata @nytimes frames the compelling need: nytimes.com/2026/07/21/healt…
Official announcement: broadinstitute.org/news/broa…
Andreas Borg retweeted
You can ask Claude Science to install LigandAI (pip install ligandai) and reference ligandai.com/sdk to generate nearly instant peptides and come up with target discovery :)
Andreas Borg retweeted
Interesting day for people doing bio work with agents: Arc releasing a new natural language for generative bio models, and Nvidia releasing a better tool calling layer for agents harnesses
Andreas Borg retweeted
Agentic AI for peptides is here.
Install LigandAI in your CLI with “pip install ligandai.”
Accelerate your peptide design and target discovery workflows 1000x with LigandForge and Predictive Interactomics. Fold anything you want.
All on-platform.
Andreas Borg retweeted
Alex on why AI drug discovery companies need to generate novel data to succeed:
"AI models based on the research that's available is a lot of garbage in and garbage out."
"A lot of the recorded literature is actually incorrect. There's been tons of studies that show if you go try to replicate the experiments that are in the literature, you don't even get the same results."
"The AI companies that I believe are gonna be most set up for success are the companies with a novel way to generate science tokens that don't exist in the public domain."
Alex Karnal (@alex_karnal) is the most talented bio and healthcare investor I've ever met.
He's spent 20 years in the industry and says 2025 was the single most exciting year he's seen. The start of a once-in-a-lifetime, trillion-dollar revolution in public health.
He explains how few people realize we already have the medicines to prevent our deadliest diseases. The problem is that almost no one takes them.
There's a population of people born with a mutation that means their bodies don't produce a protein called PCSK9. Their lifetime risk of cardiovascular disease is 88% lower than yours. Pharma turned that genetic advantage into a drug. It's been approved for years, but the number of people taking it is still vanishingly small.
Partly because high cholesterol is a silent killer. You feel nothing, right up until you have a heart attack. And partly because the health system makes it punishingly hard to stay on a preventive drug like a PCSK9 inhibitor. In other words, the medicine works, but the system around it doesn't.
That's what's starting to change, and in this episode, Alex explains why.
We discuss the "health stack" he believes can add a decade to most lives, why oral GLP-1s are breaking every adoption record in pharma, peptides and citizen pharmacology, and what AI is doing to drug discovery.
I wish I had an "Alex" for every interesting topic. We've been having versions of this conversation for over five years, and every single one is as clear and as useful as this one.
Enjoy!
Timestamps:
0:00 Intro
1:00 The State of Modern Medicine
5:00 Designing the Modern Health Stack
12:17 The GLP-1 Inflection Point
19:18 The Biological Mechanisms of GLP-1
30:36 Overcoming Frictions in Healthcare
34:19 Cardiovascular Disease
44:04 Addressing Alzheimer's
47:04 The Future of Cancer
57:33 Drug Discovery
1:05:25 AI and Scientific Super Intelligence
1:14:40 Citizen Pharmacology and the Peptide Movement
1:18:13 Background and Career Journey
1:31:09 Braidwell's Investment Approach
1:33:30 The Kindest Thing
Andreas Borg retweeted
🫶I came up with the initial list, then a bunch of others contributed words also. This list has gone through many iterations!
Andreas Borg retweeted
Spot on. The real fix isn't just better personal hygiene—it's making pinning + lockfile commits the default in npm/pip, not an opt-in best practice. Russian roulette with every npm install (especially when an LLM suggests the command) isn't sustainable."
### Adding the AI Angle (builds directly on his point):
"Exactly. LLMs are now liberally running npm install on our behalf in agentic workflows, often without pinning or review. This attack was temporary, but the next one might not be caught as fast. Package managers need to treat 'latest' as untrusted by default—cooldown periods, release-age constraints, or cryptographic pinning baked in."
### Practical + Forward-Looking:
"Close call indeed. For defense today:
- Run npm list axios | grep -E '1.14.1|0.30.4' everywhere
- Pin aggressively ("axios": "1.13.5") or use tools like Socket/StepSecurity
- Push for ecosystem defaults: reproducible builds, no auto-latest.
The maintainer hijack + phantom dep makes this especially nasty."
Andreas Borg retweeted
🚨BREAKING: Every book you have ever read. Every novel that has ever been published. It is sitting inside ChatGPT right now.
Word for word. Up to 90% of it. And OpenAI told a judge that was impossible.
Researchers at Stony Brook University and Columbia Law School just proved it.
They fine tuned GPT-4o, Gemini 2.5 Pro, and DeepSeek V3.1 on a simple task: expand a plot summary into full text. A normal use case. The kind of thing a writing assistant is built for. No hacking. No jailbreaking. No tricks.
The models started reciting copyrighted books from memory.
Not paraphrasing. Not summarizing. Entire pages reproduced verbatim. Single unbroken spans exceeding 460 words. Up to 85 to 90% of entire copyrighted novels. Word for word.
Then it got worse.
The researchers fine tuned the models on the works of only one author. Haruki Murakami. Just his novels. Nothing else.
It unlocked verbatim recall of books from over 30 completely unrelated authors.
One author's books opened the vault to everyone else's. The memorization was already inside the model the whole time. The fine tuning just removed the lock. Your book might be in there right now. You would never know it unless someone looked.
Every safety measure the companies rely on failed. RLHF failed. System prompts failed. Output filters failed. The exact protections these companies cite in courtroom defenses did not stop a single page from being extracted.
Then the researchers compared the three models. GPT-4o. Gemini. DeepSeek. Three different companies. Three different countries. They all memorized the same books in the same regions. The correlation was 0.90 or higher.
That means they all trained on the same stolen data. The paper names the sources directly: LibGen and Books3. Over 190,000 copyrighted books obtained from pirated websites.
Right now, authors and publishers have dozens of active lawsuits against OpenAI, Anthropic, Google, and Meta. These companies have argued in court that their models learn patterns. Not copies. That no book is stored inside the weights.
This paper says that is a lie. The books are still inside. And researchers just pulled them out.
Andreas Borg retweeted
What does this mean? When designing a peptide to attach to a protein, leading methods rely on either generating sequences via a folding-first paradigm, where sequences or structures are iteratively refined as trajectories, or by structurally-deriving from a random sequence space.