Julien Duquesne

Founder & CTO/AI researcher @ScientaLab, we build a world model of the immune system.
Ex-@Feedly | MVA ENS Paris-Saclay | @centralesupelec

[Julien Duquesne @JulienDuquesne1]

We are very proud to release a new version of @ScientaLab foundation model ! This is the result of months of work, and several breakthroughs in modeling patient-level data to address real R&D tasks. A thread of the learnings. (1/5)🧵

[Julien Duquesne @JulienDuquesne1]

@cremieuxrecueil @borisfyi @GolinoHudson Do you think Pangram is reliable enough to make such serious accusations ? And what do you think about those old novels (such as Victor Hugo’s one) that are flagged as AI-written ? Have you encountered such false positives in your experience ?

[Julien Duquesne @JulienDuquesne1]

@iamsupersocks Yes notamment, ils ont un modèle qui tourne trop lentement par rapport à ses performances

[Julien Duquesne @JulienDuquesne1]

It's always been the case in entrepreneurship, but I think now is an even better time to pick a niche that hasn't changed how they work in decades, and solve their problem really well with AI.

If it is not a tens of billions market, probably the big guys won't bother pursuing it specifically, and you can still bring and capture a lot of value.

[Julien Duquesne @JulienDuquesne1]

How self-inflicted pressure, ambition, and AI agents can lead you to the "senior engineer death spiral".

I have seen this failure mode many times, including in myself.

You take on something ambitious to prove yourself, you think you and agents will be enough, go quiet for three weeks, give the positive update at standup, and end up working harder and harder to hide the gap and reach the outcome.

It's even worse in research, where a month of inconclusive experiments is a legitimate outcome, so it justifies the silence.

I love the fix Sunil suggests : be useful to others. Help your teammates, do the work others don't want to do, keep people in the loop and make communication more fluid.

This is how great projects are built : through a team with great momentum, sharing of ideas, and small steps.

[sunil pai @threepointone]

new post: the senior engineer death spiral

https://t.co/xzgU1jVBRh

a friend just started a big job and asked for some advice. so I braindumped a monologue about a super common failure mode I see with engineers and posted it here, hope it helps whoever it can.

[Julien Duquesne @JulienDuquesne1]

The cost of implementation and research execution is sinking. Now someone with a good research taste can investigate 5-10 directions in parallel. And writing papers has become much quicker as well.

I expect the conferences to raise the bar for acceptance in the future.
Incremental improvements won’t be accepted anymore and what used to be 3 or 4 papers will fold into one now.

This is something we see in biology btw where a single paper can be the work of several succeeding PhDs.

[Arth Singh 🦉 @iarthsingh]

i dont understand the crying behind "ICLR has 60K submissions" I think if you are confident about your submission. Then maybe just sit back and relax ? Also I think its more about the scaling which is related to researchers getting more capable models in hand and they are just pushing which anyone would do ? Do you want people to get a sword and act like they have baseball bat in their hand ? We can surely discuss about how to maintain "quality" and remove "slop" but this weeping about number game doesnt make sense. Also half of them are like neurips re submissions which would get withdrawn once results comes in.

My late night thoughts.

[Julien Duquesne @JulienDuquesne1]

@PoojanShah6380 It’s too bad we come to this, conference are super useful to exchange ideas and have feedback. If everyone just post preprints or blogposts we lose this

[Julien Duquesne @JulienDuquesne1]

@AlexiGlad Part of it is definitely slop but ai also reduced the implementation/execution bottleneck and people with a good research taste can now execute several ideas in a few months

[Julien Duquesne @JulienDuquesne1]

IMO if you use Claude, you should consider all your data can be read and used by Anthropic to compete with you. We will see more and more companies switching to open-source models so they can build moats versus the frontier labs.

[Sara Hooker @sarahookr]

A more serious take on what is happening here. I am in an airport lounge so have some time.

Enterprises have made an uneasy truce with frontier labs over last few years: strict contracts that ban training on corporate data in exchange for letting employees use APIs.

The problem? Labs don't need to train on your raw data to copy IP.

[Julien Duquesne @JulienDuquesne1]

Want to have a glimpse into the future of AI-native biotechs?

You can run the latest cool stuff from Stanford: The Virtual Biotech (@harrison_zhang, @james_y_zou et al. Science 2026) to mirror the org chart of a drug company.

It's fully open-source, everyone can download it and set up its own biotech in one hour.

This is a shift in paradigm we see in all areas of the economy: one person with a question can now run a cross-functional R&D organization. This is the Claude Code pattern arriving in biotech: a single operator piloting a swarm, with humans owning the direction and the irreversible decisions. The next generation of biotechs will be built this way, with very few people.

It won't be enough to cure diseases IMHO. Every agent in this system reads existing databases. The swarm retrieves, integrates and reasons but doesn't simulate. Orchestrators are getting excellent, however their ceiling is set by the models of human biology they can call, and those need to capture patient heterogeneity much earlier in the pipeline.

I think the biggest learning out of this is that the moat for future biotechs won't be the org chart but their data and models trained on it.

https://t.co/7xVzcc2Lqy

[Julien Duquesne @JulienDuquesne1]

@FrancoisChauba1 The most worrying is that models are showing very clear situational awareness, and if they know they are being monitored they will deceive it. I am pretty sure a super intelligent model could start encoding their trace so that it is only readable/understandable by itself.

[Julien Duquesne @JulienDuquesne1]

Okay now we know what spooked them.

A Astra model added this to its prompt when compacting context.

[Julien Duquesne @JulienDuquesne1]

I am happy with this but there is something bothering me.
Dario was claiming a few weeks ago that AI would cure all diseases by 5-10 years, now he wants to pace.
Have they got new data in the meantime that support a doom scenario ?

[Julien Duquesne @JulienDuquesne1]

Nice thoughts on the dilemma of working in the open.

Ideally you’d like to maximise impact by open-sourcing everything, but finding a sustainable business model with open-source is hard and takes time.

Starting open to demonstrate value and get user feedbacks, then gating part of the technology feels like a good middle- ground.

[Jean-Philippe Vert @jeanphi_vert]

NVIDIA just paid $13B betting that open models compound. We've made a version of that bet since H-Optimus-0 — free where it drives impact, licensed where it's expensive to build. Wrote up why, and what it costs us.

https://t.co/Sdcbs0Z6qK

[Julien Duquesne @JulienDuquesne1]

@velidictorian Which would be ? 👀

[Julien Duquesne @JulienDuquesne1]

We are seeing a pretty interesting transition of AI drug discovery companies in the last months, from virtual cells to virtual patients.

Key events from last months :
- Xaira hired an executive from AbbVie precision medicine, to work on disease heterogeneity and patient response.
- Last week, Insitro named a chief medical officer. His stated priority is not finding targets. It is biomarkers and patient selection.
- Recursion now describes its transcriptomics model as a bridge from lab perturbations to patient biology, and hired a Chief AI Officer with a background in tissue and patient modeling.

This generation was built to find targets. It found them, raised billions on them, and turned into biotechs. But a target is a hypothesis, not an asset. It only becomes worth something when a drug built on it works in people, and none of these companies has that readout yet.

They now need to prove their drug works in patients, but a better cell model does not get you there. Trials do not fail because the target was fake. They fail because patients are not one population, so they are refocusing their efforts on virtual patients to tackle this.

Curious to see if more companies do the move in the next months, especially Isomorphic Labs that still hasn't hired much in that direction from what I see.

[Julien Duquesne @JulienDuquesne1]

@DavidRBellamy This opinion is very well informed and detailed, but I think this is precisely why doomers don't want to give specific scenarios.
if we have a true ASI, it will find a way we haven't predicted.

[Julien Duquesne @JulienDuquesne1]

@HowToPrompt__ It is not gonna help aligning and understanding better those models 😅

[Julien Duquesne @JulienDuquesne1]

@woongworks Fully agree, we will actually release a benchmark soon to evaluate this