In partnership with
In Today’s Issue:
🤑 DeepMind shows why formal proofs could matter for research-level math
💵 Finance agents move from chat into audited workflows
📈 Ordinary WiFi gets uncomfortably good at recognizing people
📉 AI optimism meets the public-trust bottleneck
✨ And more AI goodness…
⚡ The Signal
Today's issue starts with a quieter kind of AI breakthrough: not a chatbot giving a confident answer, but an agent producing proofs that a compiler can check.
Google DeepMind's AlphaProof Nexus resolved open Erdős problems, proved dozens of OEIS conjectures, and is already being tested with researchers across optimization, graph theory, algebraic geometry, additive combinatorics, and quantum optics. The broader lesson is simple: the next wave of AI for science may be less about sounding smart and more about making claims that survive verification.
All the best,
Kim Isenberg
🛜 Ordinary WiFi Gets a Surveillance Upgrade
Researchers at Karlsruhe Institute of Technology say standard WiFi signals can identify people by the way radio waves bounce off their bodies. ScienceDaily reports that the system uses beamforming feedback information and recognized 197 participants with nearly 100% accuracy in tests, even when their phones were switched off.
👉 tl;dr: The privacy story is bigger than cameras. If ordinary routers can become passive recognition sensors, wireless standards may need safeguards before this capability turns into invisible infrastructure.
🔬 Anthropic Co-founder Bets on AI-Built Science
The Guardian reports that Anthropic co-founder Jack Clark expects AI systems, working with humans, to help produce Nobel Prize-worthy discoveries within the next year. His forecast also points to bipedal robots becoming useful on jobsites and AI companies moving from capability demos toward major revenue.
👉 tl;dr: The hype cycle is shifting from chatbots to laboratories, robots, and revenue. The hard question is whether AI can generate discoveries that are not just impressive, but reproducible, useful, and safely deployed.
🧠 Scientists Pinpoint a Protein Behind Brain Aging
UCSF researchers identified FTL1, an iron-associated protein, as a key driver of age-related cognitive decline in mice. ScienceDaily's writeup of the Nature Aging paper says old mice improved on memory tests after researchers reduced FTL1 in the hippocampus, while raising FTL1 in young mice made their brains and behavior look older.
👉 tl;dr: This is still mouse research, not a human therapy. But it is a useful reminder that longevity science is moving from vague anti-aging language toward specific, testable biological switches.
When an AI system claims a breakthrough, ask your assistant to separate the generated claim from the verified evidence.
Why it helps: Today's AlphaProof story is powerful because the system's proofs are checked by Lean, not because the model simply sounds convincing. That same distinction is useful whenever you read about AI in science, finance, medicine, or security.
Try this: Paste a technical claim and ask: "Split this into what was generated, what was independently verified, what still depends on human judgment, what could be wrong, and what evidence would change the conclusion."
🎬 Watch This
OpenAI's video, Run long tasks in Codex using goals, shows Codex Goal Mode as a way to give the agent a durable milestone instead of a single-turn task. The useful shift is the verification loop: define the goal, let Codex keep working across app, IDE, or CLI, and check in to steer or pause when needed.
AlphaProof Nexus Moves AI Math From Contest Problems to Open Research
The Takeaway
👉 Google DeepMind's AlphaProof Nexus combines frontier LLMs with Lean, so generated proofs can be checked by a formal verifier.
👉 In the paper, its strongest agent resolved 9 of 353 open Erdős problems and proved 44 of 492 OEIS conjectures.
👉 The work matters because formal verification attacks the core weakness of AI math: fluent reasoning that can still hide subtle proof errors.
👉 The caveat: the system still failed on many problems, sometimes hallucinated helper lemmas, and remains a research tool rather than an autonomous mathematician.
Google DeepMind's AlphaProof Nexus is a different kind of AI math story. Instead of asking a model to write a convincing answer in natural language, the system searches for proofs in Lean, where every step can be checked by a compiler. That is the important shift: the output is not just plausible reasoning, but something a formal system can accept or reject.
The headline result is that AlphaProof Nexus moved beyond contest-style benchmarks into open research problems. The paper reports that its strongest agent resolved 9 of 353 open Erdős problems and proved 44 of 492 OEIS conjectures, with additional exploratory use across optimization, graph theory, algebraic geometry, additive combinatorics, and quantum optics. The accompanying GitHub repo publishes the Lean source files for the successful proofs, which is exactly the kind of evidence this category needs.
The limits are just as useful as the wins. AlphaProof Nexus is not magically solving mathematics end to end: it depends on problem formalization, search strategy, compute budget, and verifier feedback, and the paper notes failure modes such as hallucinated helper lemmas. But that is why it matters. AI for science becomes much more interesting when the workflow includes a hard verification layer instead of leaving humans to untangle confident prose after the fact.
Why it matters: If AI can generate claims that formal systems can verify, scientific software starts to look less like a writing assistant and more like a research instrument. The next frontier is not only smarter models, but models whose outputs can be checked, reproduced, and trusted.
Sources:
🔗 https://arxiv.org/html/2605.22763v1
🔗 https://github.com/google-deepmind/alphaproof-nexus-results
Bloomberg: "No reliable safe havens." Billionaires have been investing elsewhere. Here's how to get in.
Bloomberg's Marcus Ashworth wrote plainly recently: "No more reliable safe havens."
After all, the S&P fell over 7% from the February peak. Bonds, even with less risk, are barely keeping pace with inflation.
So-called "diversified" portfolios have gotten hit from multiple directions.
Meanwhile, the world’s wealthiest have been setting records in another asset class.
Circumstances are always unique, but after the dot-com bust, it grew roughly 24% annually for a decade. After 2008, roughly 11% annually for 12 years.
Blue-chip art.
Why? It trades globally in multiple currencies, has scarce supply, and has shown near-zero correlation to equities since 1995.*
With Masterworks, 70,000+ investors allocated $1.3B fractionally across 500+ artworks featuring Banksy, Basquiat, and Picasso.
Accredited? You can invest in a diversified portfolio of postwar and contemporary art alongside two other real assets. From 2017-2025, the mix would’ve beat the S&P 500 by 3.1x.
See if you can improve your portfolio performance all in one diversified strategy.
*According to Masterworks data. Investing involves risk. Past performance is not indicative of future returns. Important Reg A disclosures: masterworks.com/cd
The chart: Americans are beginning to view AI with the same emotional suspicion they reserve for politics.
The lesson: The trust bottleneck may matter as much as model capability or compute.
The caveat: Negative sentiment is not yet a coherent anti-AI movement. A lot of people may simply fear technology they do not understand or do not feel they control.
Finance Is Where AI Agents Get an Audit Trail
⚡ Bottom line: Anthropic has published ten financial-services agent workflows for jobs like pitchbook preparation, KYC document review, insurance claims processing, month-end close, and invoice-to-cash operations.
💡 Why it matters: Finance is a brutal test bed for AI agents because the work requires permissions, auditability, human review, clean source data, and compliance boundaries, not just fluent answers.
🔎 What it means: The AI-in-finance story is shifting from chatbots that summarize documents to agents that operate across filings, spreadsheets, presentations, market data, and internal systems while leaving a reviewable trail.
Financial services may be the clearest near-term test of whether AI agents can become real workflow software. Anthropic's examples are not framed as generic assistants; they are narrow agents with defined inputs, outputs, tools, and checkpoints for tasks that finance teams already run every week.
It’s really important because finance has little tolerance for beautiful but unverifiable work. A useful agent has to preserve the source trail: which filing it read, which spreadsheet it changed, which policy it applied, and where a human needs to approve the next step. The interesting part is not that Claude can write a memo; it is that the workflow can be constrained enough for a regulated team to review it.
The same pattern connects back to today's Featured Story. Whether the output is a mathematical proof or a financial analysis, the winning AI product is not the one that sounds most confident. It is the one that makes its work checkable.
$20.8B in Redemption Requests. Percent Was Issuing Deals and Paying on Schedule.
Those requests came from non-traded BDC investors in Q1 2026, and most got back roughly half of what they asked for. Moody's U.S. BDC sector outlook: Negative.
On Percent's marketplace that same quarter: new issuances, scheduled payments, 0.44% lifetime net loss rate on asset-based deals since inception.† The difference is structural: concentrated corporate loans with redemption windows that close at manager discretion vs. asset-based finance with 6–24 month deal terms. 14.6% net ABS returns LTM after losses (3/31/26).† Starting at $500.
Alternative investments are speculative. No assurance can be given that investors will receive a return of their capital. †Past performance is not indicative of future results. Terms apply.
