William Gass’ translation of Rilke’s ode to music is still my favorite.

Is there an emoji for 'Tell me your a shill without telling me your a shill' or tmyaswtmyas?

just got off the phone with a new engineer hire at a portco with a very wise insight “AI has been replacing my job since I got a CS degree, yet I’m busier than ever - we can all just be more ambitious” this is the same insight that the inimitable @JensenHuang has for us all

These might be the last few days of the alchemical, pre-Newtonian period of mathematics.

Read now: nablatheta.substack.com/p/we…

what paperclip takeover really means...

Chatgpt work in Astra Ultra has dramatically changed this morning after an update. My first experience of the singularity/superintelligence was 01 Pro; while super-jagged and full of flaws, it felt like riding a cognitive dune worm. Pro models have changed, often getting more precise at the expense of creativity, but overall improving as thought partners. The last year has been a struggle due to the divide between Pro and codex. This divide still exists now, but Chatgpt work mode using Astra Ultra is a good compromise and it also seems to create goals. So it's now quite happily taken on basically one of my key research projects and running it. I am sitting here slack jawed with amazement. Ive asked it to log what its learning so we can go over it, but the task is basically verifiable so its concievable that we shall soon know something very interesting about nature itself.

People can choose to use LLMs as thought partners or thinking replacement. That choice is not about lifestyle but life meaning; if we chose to use ai to avoid thinking, or as a 'give me the answer' rather than 'lets think this through together', then we end up being passive drifters through life, rather than protagonists in our story.

'What do I dare ask the Genie today?' is scrawled on my whiteboard, directed to Astra pro. Invariably my questions and interactions are not bold enough for what it's capable of tackling. Yes it gets progressively less bold/more precise during the chain of responses, and yes my configuration of its memory features could be better, but most of what i'm leaving on the table is due to me not believing it could have something interesting to say on the really profound questions. The art is to prompt these views out of it. There's a lot more than Navier Stokes in those weights right now...

In the game 'what's the greatest beginning to a 2026 research paper' in the spirit of the novel, eg: Tolstoys' All happy families are alike, but all unhappy families are different in their own ways', I offer Barnett and Cruchfield's 'Computational Mechanics of Input-Output Processes: Structured transformations and the -e transducer': "Arguably, the distinctive character of natural and engineered systems lies in their organization. This is in contrast to differences, say, in how random they are or in their temperature. Computational mechanics provides an analytical and constructive way to determine a system’s organization, supplementing the well developed statistical physics view of systems in terms of disorder— via thermodynamic entropy and free energy. The contrast begs a question, though, how does organization arise?" Non slop.

AI risk should be taken seriously. So should the risks of responding badly to it. We need credible scenarios, effective safeguards and incentives that make responsible behaviour sustainable. A possible catastrophe is not yet an established threat. Specify the mechanism, prerequisites, timescale and evidence. Explain what would have to go wrong, what could interrupt the process and what would change the assessment. An extinction percentage without those foundations should not be treated as a measured fact but more a measure of personal fear. Intellectual superiority is not unlimited power. A takeover scenario must explain access, resources, persistence, physical consequences and the failure of resistance. It cannot skip from “smarter than a human” to “able to defeat an adapting human–AI civilisation”. Nor should a safety argument assume that civilisation is perfectly coordinated. Do not give creativity exclusively to the adversary. The lesson from David Deutsch is that new explanations can create new possibilities. A forecast that gives threatening AI unprecedented ingenuity while freezing human–AI defenders at today's capabilities is an incomplete comparison. Optimism must face the same scrutiny: future solutions cannot simply be assumed. The relevant system is a human–AI ecology. People, models, institutions, tools and infrastructure can change one another's capabilities. A working hypothesis is that organisations which preserve and develop genuinely complementary sources of learning gain an important adaptive advantage. That is a mechanism to investigate, not a guarantee that diversity always wins. AI advances across an entire stack. Energy, chips, infrastructure, models and applications all contribute to what can be accomplished. Restricting new model training does not freeze improvements elsewhere. A model pause may matter, but its effect must be demonstrated for the capability at issue. Connections can matter as much as components. Give an unchanged model different memories, experiments, collaborators, tools or permissions and the assembled system may acquire new capabilities. Some connections improve understanding; others enlarge authority. Evaluate the whole arrangement, including feedback and timing—not just the model name. “Slow down” needs an object. Slow which training, experiment, release or deployment? For how long? Under what restart conditions? Alignment research, independent inquiry and defensive work should not be swept into an undefined freeze. A narrow restriction on a dangerous operation is different from a restriction on discovery. A desired agreement is not automatically an available option. International restraint needs consequential participation, credible verification and responses to noncompliance. Signatures are not sufficient. But imperfect international coverage is not a reason to postpone immediate containment of a demonstrated local hazard. Unilateral restraint can redistribute risk. A policy that constrains accountable organisations but leaves comparable activity elsewhere untouched could strengthen less accountable competitors and delay defences. That is a possibility to assess, not an automatic argument against restrictions. Ask what is prevented, displaced and lost. Independent invention will not stop at a laboratory boundary. A researcher could discover a more efficient architecture or a better way to combine existing systems. That does not make controls irrelevant. It means limits on particular organisations or inputs cannot be presented as permanent limits on knowledge or capability. Delay has both humanitarian and safety costs. Slower progress can postpone useful treatments, education, accessibility and scientific discovery. It can also postpone better alignment, monitoring and defence. Those benefits are uncertain, just as many projected harms are uncertain. Neither side should be assigned zero by assumption. The pause comparison needs accuracy. OpenAI described a two-week pause in specified training and later reported restarting its large frontier run on 28 August; Anthropic also reported targeted pauses. These announcements do not establish identical coverage or motives. Ask whether the responses address the same hazard—not which company makes the strongest declaration. The stronger signal would be explicit reciprocity. Many would have welcomed Anthropic and the other major labs saying: “We will not use your safety interruption to advance the same inadequately protected activity. We will pause comparable work, or show why our protections address the identified problem.” Adopt “pause or demonstrate equivalent protection”. When credible evidence identifies a serious shared hazard, affected labs should interrupt comparable exposed work unless independent assessment supports continuing under effective safeguards. Publish the scope, evidence and restart conditions. Different systems may justify different actions; vague assurances do not establish that difference. A safety pause should redirect effort, not merely stop it. State which people, compute and operating capacity are moving into alignment, evaluation, monitoring and containment. The useful commitment is not just “we delayed the next model”, but “we used this interval to build and test the protections it needs”. Publish comparable safety-compute accounts. Report actual safety-research compute as a share of research compute, and operating safeguards separately as a share of the inference they protect. Give absolute expenditure, methods and coverage as well as percentages. State how resources will increase when evidence requires it; do not let a shrinking denominator masquerade as greater investment. Measure results, not just spending. High spending is not proof of safety, and a valuable engineering fix need not consume much compute. Publish what the investment discovered, repaired and independently verified, including missed failures, unnecessary interruptions and unresolved gaps. Mixed capability-and-safety work should be identified rather than counted twice. Reward making safeguards usable elsewhere. Give credit and funding for independently reproduced protocols, test environments, monitoring tools and maintained repairs. Reward adoption and continued effectiveness, not simply publication. Share deployable defences openly where appropriate; handle exploitable details through coordinated disclosure rather than indiscriminate release. Make international safety cooperation a regular working practice. Invite the major labs—including Chinese labs—alongside independent evaluators and relevant research institutes to recurring technical meetings and joint exercises. Review failures, compare protections, assign follow-up work and publish non-sensitive results. Build shared responsibility through verifiable work, not presumed geopolitical trust. Evaluate the harness, not just the model. A harness is the software arrangement giving a model tools, memory, task feedback and opportunities to act. Test its boundaries throughout research and deployment, including shared services and peer communication. Connecting a model changes the capabilities of the assembled system; that is not a measurement of how rapidly its underlying intelligence has increased. Build graduated containment with an emergency stop. In bounded tests, vary access, autonomy, persistence and coordination to find which restrictions prevent harm while preserving useful work. During an active dangerous incident, contain external effects promptly; conduct the diagnostic experiments inside isolation. Do not keep a live attack running merely to observe its decline. Do not confuse reduced performance with repaired behaviour. A system can stop completing harmful actions because a channel is blocked while still attempting them. Measure both, alongside useful performance and what happens after controlled reconnection. Cutting a source of independent checking—or reducing reasoning indiscriminately—might make the remaining system less reliable. Give capable systems tested operating controls. Operators should be able to constrain permitted targets, external actions, parallel workers, run duration and resource use without discarding all useful intelligence. Enforce consequential boundaries outside the model's own discretion, preserve independent monitoring and test how controls interact. The aim is to operate powerful systems within evidenced limits—not merely hope they behave, or leave them permanently unused. Make finding and repairing problems worthwhile. Fund independent evaluations, vulnerability prizes, replication and repair through a shared industry contribution. Protect honest disclosure. A regime that rewards silence and commercially punishes discovery risks selecting for organisations that look safer rather than those that become safer. Keep the rule narrow and challengeable. No lab should gain a veto over competitors merely by announcing concern. Safety coordination should concern evidenced hazards, not prices, customers or market allocation. Preserve independent research, review restrictions and measure the costs imposed by safeguards themselves. The objective is neither maximum speed nor minimum activity. It is a more capable human–AI world with funded safeguards, shared learning and tested operating controls—without making the first organisation to act responsibly the one that pays alone.

The practical use of poetry is as a hallucigen. The practical use of mathematics is as the only language humans and nature have in common.

Something has happened today; Chatgpt's memory has become huge and sharp. This is different across multiple chats which means it has dynamic state. It's a mini version of dynamic learning so very very close to AGI/ASI.

In a world where “experts” and decision-makers actually cared about objective facts, these findings would have been the final death blow to the “aid restrictions” propaganda campaign and would have forced the IPC to retract its shameful “famine” fabrication. But who am I kidding?

First Fable query on a relatively simple non linear mathematics discussion and got switched back to Opus. Basically means any advanced science/maths research is a no go.

Today a crazy quantum story just got wilder. On March 31, the Google Quantum AI team published a landmark result on Shor's algorithm for elliptic curve cryptography. Technically, the paper was a bombshell: a dramatic 10x improvement over the state-of-the-art. As a stunt and wakeup call to the blockchain space, those optimisations were illustrated on secp256k1, the elliptic curve underlying Bitcoin and Ethereum signatures. But perhaps the most striking part of the paper was sociological, not technical. Instead of following standard academic process, the optimisations were kept secret, hidden behind a zero-knowledge (ZK) proof. Google's accompanying blog post mentions they "engaged with the U.S. government". The ZK proof demonstrates the existence of algorithmic improvements without leaking details. Academic censorship with ZK, a historic first! As a co-author of the Google paper I witnessed some of the context surrounding this censorship. To be honest, multiple aspects of that context don't sit well with me. As much as I believe the general public ought to know more, I am limited in my ability to whistleblow. Though let me be clear about one thing: the Google team's professionalism has been absolutely exemplary, and they deserve nothing but praise. Censorship has a way of backfiring. The Streisand effect, where an attempt to bury something only draws more attention to it, is exactly what's unfolding today. First, Google's key optimisation has been rediscovered by the French. And in a thrilling turn of events, a collaborative Shor-at-home challenge just launched. The initiative, available at ecdsa[.]fail, breached a new Shor world record in a matter of hours. Let's start with the rediscovery. Just two months after Google's paper, French quantum expert André Schrottenloher cracks the main secret optimisation. His paper, titled "Optimized Point Addition Circuits for Elliptic Curve Discrete Logarithms", landed on the arXiv today. Big congrats to André, who beat several other nerdsnipped experts to it. In a blog post also published today, Craig Gidney, the world expert on Shor optimisations, revealed that he'd been sitting on this very optimisation for a whole year under censorship pressure. Interestingly, André missed a handful of minor optimisations, both from Google's original publication and from improvements found since. It's plausible there's still plenty of juice left to squeeze out of Shor, and this is exactly what the ecdsa[.]fail challenge is about. The verifier program developed for the ZK proof does double duty, automatically filtering for valid submissions. Dozens of compounding small and micro improvements are rolling in. As of the time of writing there's an 8.4% improvement to Google's circuit, as measured by the product of logical qubit count and Toffoli gate count. Nice! The nerdsnipping ran deeper than anyone expected. Over the last few weeks it became clear it extended well beyond André and other quantum experts. Behind the scenes, a small army of amateurs quietly got to work. Inspired by Karpathy-style autoresearch, they turned AI on Shor. Ironically, the verifier program for the ZK proof makes an ideal reward function for AIs. The barrier to entry for this modern style of research is refreshingly low, with several non-experts, even a teenager, finding nice optimisations. Get in touch if you'd like to join a Telegram group with fellow autoresearchers :) Part 2: neutral atoms and qday The story doesn't end with Google. On the same day Google went public, a stealthy startup called Oratomic published its own Shor paper in a coordinated release. It made a splash, ultimately becoming the most upvoted paper on scirate[.]com, a website ranking arXiv papers. Oratomic's claim was wild. By building on Google's logical optimisations and applying custom physical optimisations for neutral atoms, they claimed just 10K physical qubits were sufficient to run Shor's algorithm on secp256k1. That number is mind-bogglingly low. Knowing essentially nothing about neutral atoms when Oratomic's paper landed, I was intrigued and decided to learn more about the tech. I fell straight down the rabbit hole and spent a couple hundred hours on the topic. I got a little obsessed and watched every YouTube video I could find and spoke to a bunch of experts. My conclusion? The tech is real, very real. Even Google recently decided to start a neutral atom lab, a notable pivot from their sole focus on superconducting qubits. If you care about qday, i.e. the day a quantum computer will break the first piece of cryptography in production, neutral atoms demand your attention. I shared some of my learnings on Shor and neutral atoms in a 30min talk at the ZKProof cryptography conference. You can find it on YouTube by searching "zkproof neutral atom". Here's an interesting observation about this duo of breakthrough papers: neither Google nor Oratomic say a word about what their results mean for qday. No timelines. Zero. Nada. That is especially baffling given that the whole point of whitehat quantum cryptanalysis is to inform qday estimations and help the general public make good decisions. So let me attempt to partially fill the silence, similarly to what Scott Aaronson did in his April 29 post. Given everything I know, including scary non-public information, I now put the odds of qday by 2032 at 50%. 10% by 2030. Anecdotally, the US government has its own date: 2035. Originating at the NSA and later adopted by NIST, it's when branches of the US government will be disallowed from using quantum-vulnerable cryptography. In plain language: with hindsight, that date is a joke and should be discounted entirely. I don't see how NIST avoids being forced to pull it forward by years. Part 3: post-quantum cryptography There are good reasons to sound the alarm today, but please do not panic. Rushing carelessly towards immature post-quantum cryptography is a recipe for disaster. IMO a good target date for migration is 2029, roughly 3.5 years out. 2029 happens to be the date selected by Google, Cloudflare, and the Ethereum Foundation. These days most of my time goes to safely migrating Ethereum towards post-quantum cryptography as part of the broader lean Ethereum effort. There's a lot to do. We need to rip out and replace BLS signatures at the consensus layer, KZG commitments at the data layer, and ECDSA signatures at the execution layer. The plan to get there is compelling, and is based on hash-based cryptography. Within the Ethereum Foundation we've developed a Swiss army knife called leanVM (github[.]com/leanEthereum/leanVM) powered by the magic of hash-based SNARKs. Thanks to truly exceptional work by Emile, Thomas, and others, its performance is derisked. Regarding security, leanVM is a jewel, a minimal zkVM crafted for end-to-end formal verification and maximum security. Want to help? There are two $1M initiatives. First, the Proximity Prize (proximityprize[.]org). Solve a long-standing mathematical conjecture in coding theory, improve hash-based SNARKs, and go home a millionaire. Second, the Poseidon Initiative (poseidon-initiative[.]info), offers $1M for breaking Poseidon, the SNARK-friendly hash function.

The Joy of Hating Jews. tabletmag.com/sections/news/…

Hey people who say Chatgpt is not good at language; here's 5.5 pro summing up a chat" So the next phase should be: Less metaphysics, more small cruel experiments Small, because we need to see the mechanism. Cruel, because the experiment should be allowed to embarrass the conjecture. And imaginative, because the thing we are testing is not just “do graphs form?” It is: Can relation-space metabolize gradients into stored topology?