AI continues to move at breakneck speeds. The main questions we faced this week is the retention power (we hear already of one large AI co who has had so much churn/costs it's going down) and what a future with superintelligent, infinitely-parallelizable, agentic computer users looks like.
But this is also a week that saw one of the worst ever stock crashes, with many questioning if the US administration is playing 5D chess or just completely insane. Many institutionals are closing their USD longs, and just pouring more capital in Europe and China.
The speed of change seems to be pervasive, not remaining confined to technology but touching everything around us.
This seems also evident in the speed of technology window opportunities compressing more and more, with the window to create massive AI companies almost largely closed after just a few years (unless the Deepseeks of the world just plainly destroy the entire window).
Prompt wrappers
This week we've been able to read through the system prompts for v0, Manus and even Cursor. Reading through them really makes you wonder what the staying power of all of these tools will actually be.
They all seems to rely on Sonnet heavily, with really most of the heavy-lifting coming from Anthropic (we're not surprised to see them raising a Series E at $60B+) (we also hear that OpenAI is raising $40B at $300B valuation 🤯).
Most people are already switching from one tool to the other with very little loyalty. And so the age-old "will the models or the application layer capture the most value" question seems to not be fully solved at all as of yet.
Have fun and look under the hood to how these tools work!
GitHub - x1xhlol/system-prompts-and-models-of-ai-tools
Contribute to x1xhlol/system-prompts-and-models-of-ai-tools development by creating an account on GitHub.
Manus panic
Leave it to Manus (do let us know if you have an invite for us)
Multimanus: welcome to dystopia!
One demo: 14 min demo of Manus in action.
OSS Manus: https://github.com/camel-ai/owl. We already have an open source implementation of what Manus does, based on Camel AI. It seems the staying power question is even more relevant than we can imagine.
Superintelligence strategy
Dan Hendrycks, Eric Schmidt and Alexandr Wang propose a strategy around Mutual Assured AI Malfunction (MAIM, our times' MAD) for nonproliferation of weaponizable AI capabilities, and offer a framework for competitiveness.
And an article by Hendrycks and Schmidt on the nuclear-level risks of superintelligent AI.
Anthropic’s Recommendations to OSTP for the U.S. AI Action Plan
Reflection: a path to superintelligence + a piece by Sequoia
Other AI
The arms race
ReArm Europe
Ursula von der Leyen announced a $800bn, five-part plan to boost Europe's defence capabilities.
Update from a world with European fiscal packages, security funding, and US tariffs.
The S&P 500 is down 1.8% in 2025, European stocks (.STOXX) have risen 9% to a record high, and Hong Kong tech stocks (.HSTECH) surged nearly 30%; halved bullish bets on the U.S. dollar to $16 billion; pushed the euro to a four-month high above $1.07.
Shield AI raises $240 million
The defence aircraft company has raised at a $5.3 billion valuation from investors including L3Harris and Andreessen Horowitz to expand Hivemind Enterprise, an AI-driven autonomy platform accelerating development for OEMs and primes in the industrial base.
Alpine Eagle raises €10.25 million
Munich-based startup Alpine Eagle has secured seed funding from IQ Capital, General Catalyst and HCVC to advance its Sentinel counter-drone system, an airborne platform using AI-driven modular sensors and kamikaze interceptors for cost-efficient threat detection and neutralization.
Can the US switch off Europe’s weapons?
As it turns out, “Most European militaries depend heavily on the US for communications support, for electronic warfare support, and for ammunition resupply in any serious conflict.”
The future of labs
Novo Nordisk Says AI Is Reliable Enough to Produce Sensitive Documents. And two more notes from the labs of the future.
Towards an AI co-scientist
A multi-agent system built on Gemini 2.0 employs a generate-debate-evolve approach for hypothesis generation, using a multi-agent architecture and tournament evolution process. The system demonstrates enhanced biomedical discovery, validated through drug repurposing, target discovery, and bacterial evolution studies.
Lila Sciences raises $200 million to build scientific superintelligence
Lila announces a platform integrating AI with autonomous labs to conduct thousands of parallel experiments, with the lofty goal of "reinventing the scientific method". Also in the NYT.The Einstein AI Model
Thomas Wolf, co-founder of Hugging Face, discusses scientific research, genius, creativity, and move 37.
Research
A safer, better Ozempic
Stanford Medicine researchers have identified a peptide that suppresses appetite and reduces body weight in mice without side effects like nausea or muscle loss.
Cultivated autologous limbal epithelial cell (CALEC) transplantation for limbal stem cell deficiency
A phase I/II clinical trial (so feasibility and safety) evaluates cultivated autologous limbal epithelial cell transplantation to treat limbal stem cell deficiency (LSCD), demonstrating corneal regeneration and improved vision.
Emerging supersolidity in photonic-crystal polariton condensates
Researchers have transformed laser light into a supersolid state, achieved by firing a laser at a gallium arsenide semiconductor with microscopic ridges, creating polaritons - hybrid light-matter particles that self-organize into a crystalline lattice while maintaining superfluidity.
Weeping willow
A team from the USTC in Hufei, China reports that their 105-qubit Zuchongzhi 3.0 quantum processor completed a random circuit sampling task in seconds, a feat they claim would take the Frontier supercomputer 6.4 billion years, outpacing Google's Willow by a million times. All usual QC caveats apply.
Other stuff we liked
Sesame Conversational Demo: welcome to the future.
"Sing, o Muse": fiction AI writer.
Speed to proof/value
A run in Shenzhen's Talent Park
Eric Schmidt joins Relativity Space as CEO: the co was having a hard time, and Eric comes in as a white knight (and majority shareholder). Let's see how this one goes!
If you think modern code review is fine as is, turn back
Playbook to fine-tune and deploy LLMs
Buy every possible lottery ticket: there was a Texas Lottery arbitrage opportunity. In April 2023, an entity called Rook TX effectively purchased the jackpot, collecting a one-time payment of $57.8 million, by acquiring virtually all of the 25.8 million possible number combinations.
Just blink once
Let us know what you think of this format, and if you have anything else we should read!
Hello everyone, and welcome to another edition of Unruly Futures, once again by Francesco who's taking charge of these more and more while I mend a couple of broken bones of my kids — apparently, they won’t be pro skiers after all.
Once again, macro, AI and robotics continue to be the most consequential topics, so they make the bulk of the weekly reading we did, but as usual there's quite a lot of other stuff. We cut down on proper science papers, but let us know if that's something you'd like to see more of.
The New World Order
Last week, Ross Douthat quoted Henry Kissinger on Trump being "one of those figures in history who appears from time to time to mark the end of an era and to force it to give up its old pretenses". While most countries are quickly finding out what those pretenses were (but they could've just listened instead of acting surprised now), we are living the semi-mystical experience of seeing a new world order. Markets are doing their job, with European defence stocks skyrocket under the pressure of a continent that is now waking up to its lack of military capabilities (just follow Rheinmetall's rally - RHM), while their US counterparts are hesitating. In all of this, we liked reading:
Cultural vibes aren't analysis by Alex Chalmers
On explaining Europe away with vibes, what is culture and what is policy.Stripe’s 2024 Annual Letter
Insightful as always, with one final section on European prosperity, dynamism, bad policies and allocative efficiency.Anduril vs. Helsing
Something on marketing and building brand identity in defence tech.
While on the other side of the cold war, things are not slowing down
Ivonescimab v Keytruda
We wrote about it here - China’s biotech sector is surging forward, despite concerns over lax regulations and quality of clinical trial data. This week's spotlight is for Akeso, whose lung cancer drug outperformed Merck’s Keytruda (the world's top selling medicine) in trials.Beijing to Shanghai in 3 minutes
As for all those things that go fast and can strike hard, Chinese scientists claim to have developed an oblique detonation engine that can operate between Mach 6 and 16 - almost 20,000 km/h.
Another Sonnet
Updates from AI land
Claude Code
Claude released Sonnet 3.7 and Claude Code, a massive deal that has initially gone relatively quiet. CC has agentic capabilities (read: is a SWE) you have to beg to please stop building. We really enjoyed this interaction on Claude coding in minutes what would've been the result of a fairly sophisticated master's research in the old world.
Anthropic raises series E
It's $3.5 billion at $61.5 billion post, round led by Lightspeed and joined by Bessemer, General Catalyst and others. Anthropic triples its series D valuation but is still way below OpenAI (we'll see where they settle with SoftBank but it was rumoured to be $300m).
Perplexity announces Comet
We're already seeing a rise in agentic browsers, and Perplexity has announced they are joining the race. While waiting for the Browser Company and many others (although we loved using Nayla, hats off to Yannick and team).
Who will watch the watchmen?
Atla releases Selene-1, the model to evaluate models, and releases an alignment platform to steer the model according to the users' scoring parameters.
For more efficient communication
One demo everyone has seen last week shows two agents switching to fast bips to chat faster than human language. Less efficient than how agents will actually communicate in the future, but fun experiment in a post-voice uncanny valley.
Sesame: Conversational speech generation
Another demo we loved, this time from Sesame: low-latency voice generation through real-time context understanding and the use of semantic and acoustic tokens. Best AI conversation we had so far.
The Physical World
Ultra-sensitive electronics
MIT researchers have developed a scalable, chip-based terahertz amplifier that integrates a distributed radiator array to generate high-power terahertz waves without requiring bulky silicon lenses. The tech enables compact, high-resolution applications in radar, communications, and medical imaging.
RNA neoantigen vaccines prime long-lived CD8+ T cells in pancreatic cancer
Really important study exploring the RNA neoantigen vaccines' capacity to induce durable CD8+ T cell responses against tumor-specific antigens. The study demonstrates enhanced immune activation, improved tumor clearance, and prolonged protection, highlighting the potential for personalized RNA-based cancer vaccines.
Amazon announces Ocelot quantum chip
Amazon’s AWS has unveiled Ocelot, its first-generation quantum chip, developed with Caltech, with a scalable architecture that slashes quantum error correction costs by up to 90% through innovative "cat qubit" technology and bosonic error correction. Quantum hardware is gaining momentum, as discussed last week - great to see Amazon joining the competition.
A manufacturable platform for photonic quantum computing
On the same note, PsiQuantum presents a silicon photonics platform with near-perfect fidelities—99.98% for state preparation and measurement, 99.50% for quantum interference, 99.22% for two-qubit fusion, chip-to-chip interconnect fidelity of 99.72%.
Glasses for robots
Georgia Tech’s EgoMimic introduces a full-stack framework using human data from Meta's Project Aria glasses to scale robot manipulation, achieving up to 228% performance gains in tasks like object-in-bowl sorting and grocery packing by co-training on human and robot datasets.
Large biomechanics space structures
Another hardcore DARPA project. If you think the future doesn't look like the future enough, reach out here.
Wayve’s London demo
On why end-to-end models are king (now it's clear to everybody) and the streets of London will soon become much more fun.
How Xpandomers work
Fascinating video by Roche on their new sequencing technology.
Phenformer - a genetic language model
GSK (+ Harvard, Max Planck, Oxford) introduce a model to 100x more base pairs than existing DNA language models. Great thread explaining why it's a big deal and what does it mean for the understanding of diseases.
They're Taking our Jobs
Perplexity's CEO Aravind Srinivas has started tweeting more and more about venture capitalists as a software. We can't expect it's only going to hit consultants. But we'll keep it light this time around, and promise a longer edition on why we VCs can start freaking out.
General Catalyst’s IPO
Despite all of the above, GC is considering an IPO. Not unsurprising given how wide the firm has gone in recent years, from hospitals to debt financing. It would be the first US fund to go public. A good excuse to look into their returns.
Other interesting stuff
Deepseek’s Revenues
The Craftsman
Fall in love and happily marry
Open-source stellarators
A guide to Cursor
Some takeaways from the AI Engineer Summit
UK battery start-up backed by Britishvolt investors aims for £1bn gigafactory
The end of YC
We're still trying it: GPT-4.5
Let us know what you think of this format, and if you have anything else we should read!
Unruly Futures 7: The Majorana Case
LA Edition + We're hiring!
Greetings from sunny Venice Beach, LA! I'm spending the week here given all of the LA funds are hosting AGMs this week, coupled with the awesome Upfront Summit (for which I don't have a ticket, so if you happen to renounce your spot..).
An important announcement: we're hiring!
It's time for us to add a 5th member of the team, which will be our in-house full-time AI hacker. We're looking for someone that has been building with AI for the past couple of years, and is interested in seeing what widely available super intelligence means for venture investing.
Check out the role description here, we will have generous referral rewards if you send us the perfect candidate!
Thanks to Francesco for writing this edition while I travel!
The Majorana Case
The disappearance of Italian physicist Ettore Majorana is one of those stories somewhere deeper down in the history iceberg meme, with theories ranging from him fleeing to Venezuela to a conspiracy to kidnap him for his works on the atom (the late '30s were quite the years). Majorana took a ferry from Palermo to Naples in 1938, and was never seen again. He was 31. Majorana left behind foundational works on neutrino physics, theory of the nucleus and mathematical physics. Fermi remembered him as a genius of the level of Newton and Galilei. This week, Microsoft contributed with their share of mystery announcing Majorana I, a quantum processor based on a breakthrough in topoconductors. The key idea is that information is encoded in Majorana Zero Modes, which ensure stability but also have a few challenges that Microsoft claims to have successfully solved. Many have reacted with skepticism, with the TLDR of the last few days' debate being Scott Aaronson's 10 FAQs. FAQ #4 pretty much sums it up: "Did Microsoft create the first topological qubit?" - "Well, they say they did!" Anyways, the Majorana case is open again. Here are some readings to dig into it.
Microsoft's Paper on Nature (2025)
One Older Paper on Topological Superconductivity and MZMs (2023)
One Explanation on Topological QC with Majorana Fermions (2011)
Some Earlier Doubts on Topological Qubits Decoherence (2012)
As for the rest of this newsletter: more AI panic, more humanoids, Arc Institute's EVO 2, AI-scientists doing pretty well, and more. Some geopolitics too, but there's so much going on that we chose to this one lean and keep all the panic for next week.
The AGI Couch
Demis Hassabis and Dario Amodei share a tiny couch to discuss AGI timelines.
Google Research publishes Titans, redefining AI memory by combining attention's precise short-term focus with a neural module that efficiently recalls long-term context. Titans outperform Transformers and modern recurrent models, setting a new benchmark for long-context AI.
ZeroBench reveals current large multimodal models struggle with complex visual reasoning, scoring 0% on a new benchmark designed to test spatial cognition. With a few interesting examples to prove yourself you're human (just in case).
Meta's new team within its Reality Labs division will be one to keep an eye on. Those hands we saw back in October will have to go on some humanoids after all (or maybe...?).
As AI turns software into services, agents marketplaces will be the new UpWork. What an economy we're heading into.
Agents for Software and Robots
Microsoft's Magma model seems a new leap toward real multimodality with real-world acting capabilities.
Some advice on pricing, partnerships and PMF, for founders building in AI-enabled services.
Scandinavian-design homes with Brian Eno playlists will need their own humanoid - 1X will offer the one that suits them best.
Figure showcases their new Vision-Language-Action model, and offer some insights into what's needed for real zero-shot humanoid control.
How Stripe's Model Context Protocol plugin for Cline simplifies your payments infrastructure. A glimpse into how most of these tasks will be performed soon.
Claude 3.7 Sonnet and Claude Code Anthropic
Last but not least, Anthropic released 3.7 and Claude Code, and gained the de-facto coding king crown. There is nothing as good, and we expect all coding model wrappers to be using 3.7 soon. There will be a few casualties for ones that don't offer the rest of the infrastructure as part of their offering. Nice always seeing the tension between the model improving and your product getting better vs being swallowed alive by it.
Bio-Tech-Bio
The Arc Institute presents EVO 2, a new AI model that can read, interpret and design entire genomes. Built on StripedHyena 2, EVO 2 has larger context window capturing wider relationships at lower costs. One interesting use is in the detection of clinically significant, pathogenic BRCA1 variants that are more likely to cause cancer. Here are some sources to look into EVO 2 and dream of the future of biology.
The Arc Institute + NVIDIA Paper
The Architecture EVO 2 was Trained on
And a few more things from the bio + tech world.
Google announces an AI co-scientist "designed to generate novel research hypotheses, a detailed research overview, and experimental protocols" and tests it on repurposing drugs for acute myeloid leukemia (spoiler: it does well). Labs producing great research with just one PI and their co-scientists: coming soon.
The benchmarking we like: how an AI co-scientist successfully rediscovered a novel gene transfer mechanism in bacteria—specifically how capsid-forming phage-inducible chromosomal islands (cf-PICIs) spread across species by hijacking diverse phage tails.
Disease diagnostics using ML of B cell and T cell receptor sequences
Immune receptors sequencing data offers a unique insight of your personal history of vaccinations and infections.
Sonal Pai makes a few predictions on what we can expect in the next few months of TechBio. Just to mention one we've seen a lot in recent dealflow: radioisotopes manufacturing.
The Fast Fast Changing World Order
Can Europe Afford to Defend Itself Without the US?
The question that is haunting European governments. The answer is yes, but the way to do it in practice is less straightforward.
Technological Progress in Chinese Political Culture
An excellent piece exploring the evolution of Chinese political culture’s approach to technological progress over the past 150 years, tracing an intellectual genealogy from the 19th-century reformers to modern leaders.
Chinese scientists have created “well-crystallised, nearly pure HD” (lonsdaleite) previously only found in asteroids, a “superior alternative” to conventional diamonds.
A Thirst for Infrastructures
A great long-ish read by Casey Handmer on a future with Starship.
VCs pondering whether (their) cash should flood data center startups as AI’s appetite grows.
Sequoia profiles RobCo in a piece on risk-taking and industrial robots.
Former Intel bigwigs nabbed $215M for a RISC-V chip venture. Worth watching.
Quantinuum’s Quantum Computer Now Fully Operational at RIKEN
Quantinuum’s “Reimei” ("dawn" in Japanese) has booted up at RIKEN, tangoing with the Fugaku supercomputer to birth a new type of hybrid brain.
Other interesting stuff
Practicalities and Psychology of Expected Value
The Diff Stock Pitch Competition
Aaron Swartz Marble Statue Unveiled at Internet Archive
Deep research tools for a fund that likes to do deep research
The Series A Fund Codex - a16z
Let us know what you think of this format, and if you have anything else we should read!