The Optimist's Case, Backed by Evidence
Packy McCormick has been making the case for optimism for nearly four years now, and this week's edition lands with unusual force — not because the tone has changed, but because the volume of genuine breakthroughs has. Seven major items, each one worthy of its own feature, crammed into a single Friday dispatch. The pace of progress isn't accelerating in theory anymore. It's happening in lab notebooks, shipyards, and semiconductor furnaces right now.
Drug Design After AlphaFold
The biggest story this week isn't a new language model. It's a drug design engine. Isomorphic Labs — the Google spinout led by Demis Hassabis, who splits his Tuesdays between the company and his day job running Google DeepMind — released a technical report on IsoDDE, its AI drug design system. The results are difficult to overstate.
McCormick writes, "On the hardest protein-ligand structures — the ones most unlike anything in its training data, where AlphaFold 3 struggled — IsoDDE more than doubles AlphaFold3's accuracy." AlphaFold's Nobel Prize-winning achievement was predicting protein structures. IsoDDE goes a critical step further: it predicts how to drug them. It outperforms AlphaFold 3 by 2.3 times on antibody-antigen modeling and Boltz-2 by nearly twenty times. It even beats FEP Plus — the gold-standard physics simulation that costs orders of magnitude more in compute time — at predicting how strongly a drug will bind to its target.
The cerebron example crystallizes why this matters. Researchers spent fifteen years believing that particular protein had only one druggable pocket. A 2026 paper later confirmed a second, hidden one. IsoDDE found both from the amino acid sequence alone, with zero hints about what molecule to look for.
"IsoDDE compresses the search phase from months of lab work to minutes of computation."
But McCormick does not pretend the problem is solved. As he puts it, "As of early 2026, no AI-discovered drug has received FDA approval. AI-designed compounds are progressing to clinical trials at roughly the same success rates as traditionally discovered ones." Biology remains unpredictably cruel once you move from a petri dish to a human body. Isomorphic Labs itself has pushed its clinical trial timeline to the end of 2026 for its first AI-designed drugs to enter human testing. The field is still in proof-of-concept.
Critics might note that compressing the search phase only creates the next bottleneck: a flood of candidates hitting a clinical trial system that was never designed for volume. Running more shots on goal only matters if the goalposts are reachable. And they have not been, for any AI-discovered compound yet.
Machines That Think Slowly
Google's Gemini 3 Deep Think mode represents something different from the usual benchmark arms race. It is a model designed to spend minutes — not milliseconds — chewing on a single problem, exploring solution paths, backtracking when they fail, and building multi-step reasoning chains before committing to an answer. The psychological framing comes from Daniel Kahneman: standard Gemini is fast and intuitive, like System 1 thinking. Deep Think is deliberate, like System 2.
The numbers are striking: 84.6 percent on ARC-AGI-2, a frontier reasoning benchmark where the next closest model scored 68.8 percent. A 3,455 Elo rating on Codeforces — which, if this were a human competitor, would place it eighth in the world. Gold medal-level results on the written portions of the 2025 International Physics and Chemistry Olympiads.
But benchmarks have been gamed before. What makes this different is the Duke University lab footage. A researcher feeds synthesis parameters into Deep Think, the model reasons through an optimized recipe for growing MoS₂ monolayer thin films — a notoriously difficult semiconductor material — and then pipes those parameters directly into lab automation software controlling the furnace, gas flows, and temperature profiles. Deep Think designed a recipe for growing thin films larger than one hundred micrometers. A precise target previous methods had struggled to hit.
McCormick writes, "It's a great time to be a researcher, and a bad time to be a problem." The model also caught a logical flaw in a proof that had survived human peer review, autonomously generated a paper on structure constants in arithmetic geometry, and solved several of Paul Erdős's unsolved conjectures from a database of seven hundred open problems.
Critics might note that catching a peer-review flaw and generating original mathematics are not the same thing — the former is error detection, the latter requires genuine novelty. And solving several out of seven hundred problems is impressive, but it is also a reminder of how far the gap still stretches.
Ships and Sclerosis
The American shipbuilding industry is, in McCormick's telling, a case study in institutional decay. China's shipbuilding capacity is two hundred and thirty-two times larger than America's. Chinese yards built over one thousand commercial vessels in 2024. The United States built eight. The American Navy is projected to shrink from two hundred and ninety-six ships to two hundred and eighty-three by 2027, as retirements outpace new construction. Thirty-seven of forty-five ships currently under construction face significant delays. The four public shipyards average seventy-six years old. The dry docks average over a century.
Then Blue Water Autonomy happened. Founded in 2024, the company unveiled the Liberty Class — a one hundred and ninety-foot autonomous steel ship with a range of over ten thousand nautical miles and more than one hundred and fifty metric tons of payload. The name deliberately echoes the Liberty Ships of World War II, which were built rapidly and at scale. The first vessel is expected to be delivered to the Navy later this year, with serial production of ten to twenty ships per year targeted.
McCormick frames it characteristically: "Every sclerotic incumbent is an opportunity for a startup to build something better, faster, and cheaper."
But then he undercuts his own optimism with a single line: "It's a good start, but we're going to need like 1,000 of those eventually to catch up." One thousand autonomous ships against a competitor that built one thousand vessels of all kinds in a single year. The gap is not just numerical. It is structural, cultural, and generational. One startup does not close it.
Chemistry Becoming Biology
The most quietly staggering story in this dispatch has nothing to do with artificial intelligence. It has to do with the origin of life itself.
Researchers at the MRC Laboratory of Molecular Biology in Cambridge — the same lab where Watson and Crick determined DNA's structure — discovered QT45: a forty-five-nucleotide ribozyme that can synthesize both its complementary strand and a copy of itself. Previous self-replicating ribozymes were one hundred and sixty-five to one hundred and eighty-nine nucleotides long, far too complex to have plausibly emerged from a primordial soup. None of them could copy themselves. Their own folded structures blocked self-replication. QT45 solves the paradox by stitching together three-letter RNA building blocks rather than adding one letter at a time. Those triplets bind strongly enough to unravel folded RNA, allowing the molecule to replicate its own structure.
As McCormick notes, "At 45 nucleotides, QT45 is small enough that the researchers argue polymerase ribozymes may be far more abundant in random RNA sequence space than anyone thought, meaning self-replication might not have required an astronomically unlikely accident. It might have been, in a sense, easy."
The triplet code that QT45 uses — three-letter RNA chunks — is the same genetic code that all life on Earth still uses to build proteins. The code itself is a still-operational fossil of the first replication system.
Critics might note that synthesizing a complementary strand in a controlled laboratory environment is not the same as spontaneous emergence in a prebiotic environment. The gap between "this molecule can copy itself under ideal conditions" and "this is how life began" remains a canyon, not a bridge. But the canyon has gotten noticeably narrower.
Bottom Line
McCormick's optimism is not naive — it is data-driven, and the data this week is genuinely staggering. But the thread running through every item is the same: breakthroughs in the lab do not automatically translate to outcomes in the world. IsoDDE compresses drug search from months to minutes, but clinical trials remain just as slow. Deep Think solves Erdős problems, but writing is still bad. Blue Water builds autonomous ships, but China still outproduces America by a factor of hundreds. The bottleneck keeps shifting. The question is whether institutions can shift fast enough to catch the breakthroughs before they stall.