After an awesome two years at DeepMind, I’m leaving tomorrow and will be working full-time on identifying and roadmapping promising new “mini-Manhattan projects”, i.e., “Focused Research Organizations (FROs)”, and trying to help them happen. Talk to me if you have one in mind!

I've met 3 researchers who are now (based on public info) involved with this still-secretive initiative. All 3 would be super FUN to do science with. One of them is a seriously top-knotch theoretical condensed matter physicist. Good very early sign on what this may actually be.

and it may explain why the synthesis is finicky

Excited to be growing the core CR team! We aim to launch 10-20+ moonshot science projects over the coming yrs, and will need exceptional people with the right mix of ‘crazy’ and ‘grounded’ to accelerate our core org + FROs across ops, talent search, etc. 5 roles open now 👇

Why MOST research should NOT be objective-driven. To get to any one objective, you have to build on a library of stepping-stones that were discovered based on subjective interestingness unrelated to that or maybe any objective. piped.video/watch?v=dKazBM3b…

Do any of y’all want to fund a Focused Research Organization

This is a cool way to build highly specific prior knowledge into neural nets: via highly specific loss functions arxiv.org/abs/1609.05566 Don’t just maximize likelihood of the data or predict the next frame—build all the constraints you can know into the loss and make NN satisfy.

What's cooler than (an awesome! wireless demo plus) promising *hypothetical* automated through-skull surgery to implant a BCI? Actual minimally-invasive surgery in humans to implant a BCI: businesswire.com/news/home/2… Why isn't the internet more excited by *endovascular* electronics!?

**** Expansion Microscopy + In-Situ Sequencing🔬🧬 **** Opening a path to what I (only semi-fancifully) call "arbitrary resolution, infinite color microscopy". With such an awesome team, who started working on this >6 years ago!

Lots that I agree with in this great essay and school of futuristic thought. Here are some highlights for me focusing on the theory of change around core science progress: 🧵

I think we are still delinquent and need more help on physics (e.g., "quantum", solid state / condensed matter, optics), astrophysics, materials science, chemistry, global health and social sciences. So if you work in those areas, and have a sense of mid-scale bottlenecks near the trunk of the tech tree and needing coordinated research programs or new tools/systems datasets to solve, please consider suggesting something using the website or emailing us about it.

Great episode. 1) In the songbird, the RL circuit used for imitation is pretty well understood pmc.ncbi.nlm.nih.gov/article… It indeed doesn't have direct supervision of motor actions token by token, but rather learns by RL to adjust its motor output to match the sound of a stored memory of a tutor song. 1.1) Note that this is an ethologically-specific, evolved reward function. "Song is learned not for external reinforcement but instead by matching vocal performance to the memory of a tutor song (Marler, 1997, 1970)." (elifesciences.org/reviewed-p…) Here is a bit about the reward function: pmc.ncbi.nlm.nih.gov/article… It has a dedicated circuit to store the tutor song memory too, so it can then use it as part of this reward function. Thus the reward function is custom designed by evolution and built up in a sequence of evolved steps -- each with specific, identifiable brain circuitry -- to enable an evolved cultural transmission process in an animal. 1.2) Evolution probably puts a lot of its programming into the reward function. That's probably why the brainstem and hypothalamus constitute most of the brain cell type diversity in the mammalian brain. nature.com/articles/s41586-0… Connectomics can reveal specific details of how this works elifesciences.org/articles/2… 1.3) Probably a lot of these are intrinsic, non-behaviorist lesswrong.com/posts/FNJF3SoN… reward functions programmed by evolution to help humans bootstrap, notably, social instincts lesswrong.com/posts/kYvbHCDe… Those social instincts can drive a bias to learn by imitation of other humans. This Ullman paper is excellent on that pnas.org/doi/10.1073/pnas.12… and inspired me to go nuts nearly a decade ago writing about how specific evolved reward/cost functions could guide animals to take on specific learning trajectories / curricula that simpler costs/objectives wouldn't bias them to arxiv.org/abs/1606.03813 2) Of course it won't be just about RL. The model used for model-based RL can be learned by self-supervised prediction, as in arxiv.org/abs/1803.10760 and in LeCun's cake diagram. There are probably smarter tricks than just backprop through a multi-layer system that the cortico-thalamic system is using to drive efficient world model learning and inference. Work from Dileep George and others takes inspiration from songbird circuits to try to model this. This model can then be used for model based RL. 3) Taking into account that AI may converge to this type of design is important for AI safety scenario planning osf.io/preprints/osf/fe36n_v…

+1 to this, don't build mirror life

we haz a new website

Think of Neuralink as the default/background state of neurotech, accelerated. They are climbing Everest with bigger team/better gear (engineering). What is really needed is a helicopter (science-intensive breakthrough).

This is a profound synthesis (no pun intended) of computation and physics/chemistry. Every molecule in the test tube can be its own sample from the generative model. Far more than one could ever deterministically synthesize. Don't completely understand it yet. Initially I didn't understand how a DNA synthesis method could possibly have enough tunable parameters to allow anything like this, but now I see the 2021 paper mentions using trimers for example biorxiv.org/content/10.1101/… So you can have clever mixtures of trimers at each step, as well as splitting and pooling to do mixtures of mixtures... which lets you do pretty complicated mixture models. Complicated mixture models can in turn approximate other models... just here in a way that is implemented physically as biased random molecular chain building rather than in a computer with random number generators and linear algebra. Probably many implications for creating and screening/pruning giant biological libraries in a ML-informed and ML-informing way. First time I've been really surprised by a biotechnology paper since magneto-genetics.

If the brain contains learning systems that can optimize cost functions / minimize error signals, what might those cost functions / error signals be? I don’t mean the obvious ones like next step prediction. Rather, the interesting ones that train a human to be a human. 🧵

Liking the Biden ARPA-C projects... #hope

Loving the new @geochurch lab website laying out value propositions for solving many of the world's biggest problems with bio-sci/tech: arep.med.harvard.edu

Congrats to @G_T_Heller and colleagues on launching Bind Research, the UK’s first FRO, on making disordered proteins druggable! bindresearch.org/fro/2025/02…

Excited to partner with @ARIA_research !

The most fun part of my job is thinking w/ smart people on bottlenecks in science and technology & how they might be solved. Don't forget to scroll down in the FROST-UK announcement to the part where we have a "Request for FROs", with some ideas for inspiration -- screenshot below:

If you can make weight-agnostic NNs, maybe you can also make backprop-unnecessary NNs: nets that can learn, with a local or reinforce-based learning rule, what would otherwise need deep exact credit assignment / full gradient to learn. cc @tyrell_turing

Hair... as a molecular ticker-tape recorder.

It’s an extremely good and important episode. But it’s a bit dichotomist about 1M reasoning geniuses in an datacenter exploding R&D via pure software, versus a much longer term unpredictable general economy wide capital deepening as the driver of R&D progress. The reasoning geniuses plus an entity at the capitalization level of a large tech company could “unhobble” key areas of R&D in a secret third way. One that looks more like orchestrating coordinated physical/bio research programs and analyzing the data from them superhumanly well and fast. To be a bit bombastic, and as one example: At the scale of a corporate research lab, and with AI as an accelerant, you could mostly check off gap-map.org/ one by one, and rapidly gain insight from the data. Make true predictive and interpretive models of cells, organisms, brains, materials and so on. Reasoners plus very good microscopes if you will. So rather than an all-software singularity, you can have a heavy-software but also heavy coordinated and centralized physical bio/chem R&D unhobbling. Which perhaps could arise mostly outside of a massively-iterative and distributed coupling with the rest of the economy at large. I’m not sure if this could get you to, say, molecular nanotech “on its own”, but as Ege says, if you had the ancient evolutionary programming of animal brains on a chip tomorrow, you might have what he considers true AGI anyway. Which is arguably a finite problem of information extraction from localized hunks of biomolecular matter, not one requiring economy wide capital deepening.

This should be in the running for most significant achievement of modern "meta-science" to date. New ARPA organization is able to define an approach to AI safety that neither traditional institutions nor the big AI labs could yet crystallize at a relevant scale.

Tractor beam. nature.com/articles/d41586-0… With citation to @Ben_Reinhardt thesis arc.aiaa.org/doi/abs/10.2514…

There are so many inspiring examples of scientific roadmaps out there: this post was our quick way of introducing new explorers of the scientific landscape to this fascinating genre. Expect more from us in terms of releasing roadmaps and thinking soon, and let us know which of your favorites we missed!

Struggling to contain my excitement about @E11BIO after seeing some new data, validation experiment plans and cost estimates for future connectomic platforms yesterday. What if we could actually know how brains, even large ones, were structured at the circuit architecture level?

2016 talk in then-student @kulesatony ‘s workshop, based on what was percolating in Ed’s lab, where I was called “Director of Scientific Architecting”. Many incredible stories and people behind this small shadow cast on the wall. Cake is still just starting to bake as we approach 2025. What would you do with a country of geniuses in a datacenter 😀?

We now know the ACTUAL translation of dog to human age -- and it is not simply "times 7". h/t @sinhanubhav cc @ArtirKel #epigenetics #iflscience cell.com/cell-systems/fullte…

Implications of this potentially huge: sciencedirect.com/science/ar…! If it works out, anyone could give organ transplant to anyone. Could also replace immune system w/ younger, cured (e.g., of a blood disease) or otherwise-enhanced donor immune system without toxic immuno-suppression.

An especially nerdy podcast appearance

Many of the most transformative ideas and people in science and technology today (and in supporting institutions) are still latent and blocked. Huge unlock and lever for positive change, led by the inimitable @tkalil2050 and team. I've worked closely with this crew, and guarantee this next step will be 🔥.

Brain computer interface FRO: I can haz ultrasound-on-chip chips? 🧠🏥 🛜 👩‍⚕️🖥️ 🔌 ℹ️ Butterfly: Yas ----🔊🔉🔉🔊🔊🔉🔊----- -----🎙️🎙️🎙️🎙️🎙️🎙️🎙️----

Turns out you can question Backprop Orthodoxy (such as it is) from within its own house: G-GLN = deep net w/ local convex loss at each neuron via gating arxiv.org/abs/2006.05964 w/ @davidmbudden @erensezener @friendly_aixi @gregdwayne, Tor Lattimore. cc @tyrell_turing @GaryMarcus

I will hold these tweets close in the dark times

"But what if you could create neural interfaces that would act as a mirror of the mind? People could see how they thought. That’s the first step in acknowledging that we really are a very flawed form of intelligence—as awesome as we are." @bryan_johnson wsj.com/articles/brain-compu…

This is something I deeply believe in and a big part of what we aim to enable @Convergent_FROs @E11BIO @ForestNeurotech

Dug up this quote on scientific roadmapping and its role in 1960s/1970s DARPA while preparing for the Convergent Research quarterly onsite aip.org/history-programs/nie…

Proud of @E11BIO's early progress here, and excited for so much more to come (that I've seen previews of as a FRO board member 😍). Optics + expansion + barcoding + AI + molecules = beginning of infinity for scalable brain circuit mapping, with wide implications across science.

Most significant and creative progress on molecular ticker-tapes since the concept was introduced around 10 years ago IMHO

🧠/acc, safe-by-design/acc, math/acc, biosecurity/acc, importantly **Pareto-topia/acc** (see link below — and therefore, nano/acc, cryo/acc, 🚀/acc), cooperation/acc, dangerous-geopolitical-conflict/decel Overall, d/acc

FAS has a short catalog of example FRO ideas up now: fas.org/initiative/focused-r… The usual boring things everyone expected: superconducting optoelectronic brains, actively scrubbing antibiotic resistance, protein-like catalysts that don’t have to fold, mapping food as medicine, …

“Reframing Superintelligence” is a must-read longitudinal.science.blog/20…

News today on systematic mid-scale science projects at @Convergent_FROs @CultivariumFRO @E11BIO and @AsteraInstitute tomorrow! Follow these and @AGamick for more.

How not to FRO

Big day in noninvasive neuro-tech: Kernel publishes on two new devices spie.org/photonics-west/pres… spie.org/photonics-west/pres… Meanwhile ARPA-like program wellcomeleap.org/1kD calls for "scale for... use in... natural environments... density of data... to model... network development"

If you like beautiful cats, it is possible you may also like this blog post running numbers for essentially all known approaches to large-scale carbon sequestration from the atmosphere: longitudinal.blog/co2-series…

Extremely excited about @ForestNeurotech's work to transform neurotechnology with minimally invasive ultrasound.

If the government got it together and ultimately put a number of brain-technology-related FROs into one centralized entity, it would look something like this, and that would be super awesome.

Increasingly thinking this is promising: vox.com/energy-and-environme…

IMHO needs to be a crash neuro program for refining & testing this hypothesis (including w/ connectomics) yesterday lesswrong.com/posts/kYvbHCDe… (That this is a blog by an ex-physicist at an institution you’ve ~not heard of = exercise to reader re how we fund neuro/psych theory.)

Facebook published a paper on a non-trivial improvement to measurement of blood flow inside strongly optically scattering biological samples, by combining signal over many optical speckles instead of just 1 effective speckle. SNR scales as Sqrt[num_pixels] research.fb.com/wp-content/u…

Nice to see @FutureHouseSF having impact on biosecurity Evals

AI bro's, if you want true long-term alpha, especially on safety alignmentforum.org/posts/ybm… you should be getting (at least) the songbird and mouse connectomes -- imagine being able to revisit this podcast but with Sutton and Dwarkesh pointing to specific brain circuits. Not only despite but also because of LLM progress.

It’s not even 10 am and I’ve tweeted about @leanprover four times

Not sure why they included me among this otherwise excellent lineup of podcast guests but this was a fun conversation!

Agreed. Likewise need virtual brains fas.org/publication/fro-meas… arxiv.org/abs/2308.06578 The low dimensionality of data that we capture from humans is also a limiter fas.org/publication/fro-scal… forestneurotech.org/ @ForestNeurotech From new kinds of fundamental data in model systems, like from molecularly annotated connectomics @E11BIO, we can probably also create more structure-to-function reveals like "it has not escaped our notice that the specific pairing we have postulated immediately suggests a possible copying mechanism for the genetic material". For example, a connectomic ground truth of this would be nice lesswrong.com/posts/H2epKysv… lesswrong.com/posts/jrewt3rL… even in animals. Fortunately, this is all addressable.

So excited to see Kernel @KernelCo colleagues announcing exciting ongoing progress on **noninvasive** human brain interfacing!!! kernel.co/hello-humanity bloombergquint.com/businessw…

Massive shoutouts to @AGamick, the first four FRO founders, @SchmidtFutures, @tkalil2050 and others for making this happen! More exciting news to be announced tomorrow as well @AsteraInstitute.

I have a list at the bottom of web.mit.edu/amarbles/www/tea…

Honored to get to work with the amazing @KernelCo team. The progress has exceeded my most optimistic expectations. I've learned enough to write more than another "book" on the principles & promise of scalable neural interfacing. Stay tuned!

For the avoidance of doubt, despite the FRO concept being formulated about 5 years after OpenAI started, @AGamick and I are happy to take full credit for OpenAI’s success, and also 99% equity in the subsidiary for-profit. Any damage to human interests caused by AGI is attributable to @SGRodriques 😜.

Go @ClariceDAiello! Fascinating as a biology/biophysics project and as a "DeSci" venture.

The beginning of something big

I wrote a series of 3 blog posts about climate change. This is the first one, other 2 coming later in Oct.

Interesting paper, cover of Nature in 2019: nature.com/articles/s41586-0… Sociological observations: -"AGI" explicitly in title -Seven co-first authors -Then, 1 author in California, rest China or Singapore. California author now at Alibaba, leading 3D chip lab damo.alibaba.com/labs/comput…

🔍 Curious about Convergent Research (@Convergent_FROs) and FROs? 🚀 Try asking our FRO GPT your questions! 🌟 chatgpt.com/g/g-Bbee1Y7ir-in… Post your most interesting response here! Disclaimer: this GPT is still in beta and is very much an experiment.

“The amount of information encoded by all of the rearranged antibody and T cell receptor genes in one person—the ‘genome’ of the adaptive immune system—exceeds the size of the human genome by more than four orders of magnitude.” Perhaps someone should do a FRO on immune-omics.

It’s always the hypothalamus

Sewing-machine insertion of flexible electrodes via small holes. Good progress on channel count + insertion, building on @m8ta's work. Also nice chip+packaging. Hard stuff. Much depends on gliosis, but Chong Xie's results promising. Laser would indeed be preferable to drill. 1/4

TIL: neurons have cilia nature.com/articles/npp20132…

larsjamesblackmore.com/Black… larsblackmore.com/losslessco… scholar.google.com/citations… piped.video/watch?v=0xoE5l… 🥂🥂🥂

The ICI is a nice antidote to the overly hardware-centric and insufficiently bio-centric thinking in most of the BCI field

People sometimes say "why haven't we solved C. elegans yet?". Arguably a gating problem was the lack of the below tool...

More on this, and other upcoming FROs and related opportunities, in a few months! cc @AGamick

A brilliant post that I think is a must-read for those interested in the AI-neuroscience intersection!

Do you like semiconductors and not like pandemics? Then this report may be for you.

essentialtechnology.blog/p/p… essentialtechnology.blog/p/i… ifp.org/mapping-the-brain-fo…

A breakthrough by these individual scientists, and also in how (computational) research is organized and funded, thus an "engine" for much more to come...

SOOO excited about @E11BIO !

What are the most interesting organisms that "synthetic biology", with its focus on E. coli and yeast, has not exploited yet?

and long on using a massive amount of similar DFT simulations to find variant materials where lower energy substitutions / easier to obtain dopings give the same effect

Michel is incredibly pioneering at both experimental and theoretical physics (obviously) — and he was also a kind, patient and inspiring undergrad thesis mentor for a lucky me. Love to see this.

To the stars!

Sam's lab is going to be 🔥🔥🔥🔥

ribo—>deoxyribo Big if true

Congrats on your paper @leanprover

Hadn't noticed this reply to famous Alberts et al "Rescuing US biomedical research from its systemic flaws" paper, from an NIH scientist: pnas.org/content/111/26/E263… "Google does not lack creativity, yet they don’t separate into 1,000 independent workers." Was it a request for FROs?

Something to feel jiggly about

TFW you've read the entire internet up to this week's issue of Nature

People are really sleeping on this lesswrong.com/posts/frApEhpy…

Fun concept: “Cosmological natural selection… of universes capable of prolific reproduction. It was originally proposed that reproduction occurs through singularities resulting from supernovae, and subsequently argued that life may facilitate the production of the singularities that become offspring universes. Here I argue technology is necessary for production of singularities by living beings, and ask whether the physics of our universe has been selected to simultaneously enable stars, intelligent life, and technology capable of creating progeny. Specific technologies appear implausibly equipped to perform tasks necessary for production of singularities…” iopscience.iop.org/article/1…

tl;dr: In work led by Dan Oran and @SGRodriques, co-PI'd with @eboyden3, and in collab w/ other important co-authors, we used hydrogels as 3D molecular breadboards to pattern materials using light, then shrunk those patterns down to the nanoscale -- a process called ImpFab.

😍

If any connectomics analysts need a side hustle

L F G !!!