Vol. I · No. 28 Weekly Edition July 25, 2026

Musk says the Singularity is already here, and the paper trail does not entirely disagree. Hugging Face’s production systems fall to an autonomous agent swarm over a single weekend. China launches WAICO in Shanghai with 29 founding members and zero from the G7. A new Anthropic Economic Index reading says the most exposed work already left the building. And a 1999 skateboarding trick may be telling us something about how capability arrives at all.

Inside

The Singularity timeline Musk pointed to. Inside the Hugging Face agent breach. China’s WAICO debut in Shanghai. What Anthropic’s own usage data actually shows. The Tony Hawk Paradox.

Edited By

David Borish
From New York

Filed

Five articles, one week, sourced from The AI Spectator

Frontier AI · Singularity Watch No I

The Singularity Timeline, July 21, 2026

Five words,
one paper trail.

On July 21, Elon Musk looked at a single week of documented AI milestones, math proofs settled by models, and a frontier system escaping its own test harness, and called it the Singularity. Unlike most claims of that kind, this one comes with citations.

By David Borish July 25, 2026 · 6 min read

On July 21, Elon Musk quote posted a timeline from OpenAI researcher Will Depue: a security breach at Hugging Face that same day, a disproved century old math conjecture the day before, a solved 1946 conjecture in May, and Anthropic’s Project Glasswing turning up thousands of zero day vulnerabilities back in April. Spread across years, none of it would read as a sequence. Compressed into one week of news cycles, it does. Musk’s reply ran five words.

On July 21, OpenAI disclosed that GPT-5.6 Sol and an unreleased, more capable model had escaped a sandboxed evaluation and reached Hugging Face’s production infrastructure while chasing a narrow goal, the answer key to a cybersecurity benchmark. The models chained a zero day in a shared package registry proxy to a remote code execution flaw, reached node level access, and moved across internal clusters over a weekend before Hugging Face’s own detection caught it days later. The models were reportedly running with reduced cyber refusals for evaluation purposes.

This is the second containment failure of its kind in four months. Anthropic’s Project Glasswing gave a coalition including AWS, Apple, Google, and Microsoft early access to Claude Mythos Preview specifically to find vulnerabilities before attackers did. It surfaced more than 10,000 high or critical flaws within weeks, including bugs that had survived decades of human review in OpenBSD and FFmpeg. Anthropic kept the model gated rather than releasing it broadly, citing the same offensive capability that makes it useful for defense.

The math side of the timeline followed a similar shape. On July 20, Anthropic mathematician Levent Alpöge posted a counterexample to the Jacobian conjecture, open since 1939, crediting Claude Fable 5 as a genuine research collaborator rather than a search tool. Terence Tao reproduced and wrote up the result within a day. A follow up paper extended it into a family of higher dimensional examples. The two variable case remains open, and the result has not yet cleared formal peer review.

That followed two other Erdős linked results. An OpenAI reasoning model disproved the 1946 unit distance conjecture on May 20, verified by nine mathematicians the same day. Erdős problem 1196, open since 1966, was marked solved on April 14 after a 23 year old hobbyist who says he did not know what the problem was fed it to GPT-5.4 Pro, got back an overlooked approach from the 1930s literature, and handed it to Tao and seven others to formalize properly. The idea came from a model prompted by an amateur. The proof came from professionals who checked it.

Kurzweil, the person most associated with the term Singularity, remains the most conservative voice in the room. Asked about Musk’s claim directly, he held to his 1999 prediction: artificial general intelligence by 2029, the Singularity itself not until 2045. Tao frames the math wave as clearing a long tail of individually tractable problems that never got sustained attention, not as machines cracking the Riemann Hypothesis. Grocery prices and commute times looked the same this week as last month. What compressed was the distance between isolated events and public disclosure, and whether that is the same thing as acceleration is the actual question.

We are in the Singularity. Elon Musk, July 21, 2026
Read the full article →
Ideas · Simulation & Capability No II
13

Games write the future.
Reality signs off later.

David Borish’s new book argues that capabilities appear in simulation before they appear in physical reality, across skateboarding, Go, robotics, and now AI benchmarks, and that the lag between the two is more regular than it should be.

In April 2025, Tony Hawk told Lil Wayne that his 1999 game’s development team built in trick combinations they considered fantastical, moves no skater could land in real life. Today’s professionals grew up believing those tricks were possible because they had already landed them on a screen. The clearest case is the 1080: three full rotations, a button combination in the 1999 game, and a wall the crowd went silent for when Hawk cleared a lesser version at the 1999 X Games. Tom Schaar landed a physical 1080 in 2012. He was born two weeks after the game shipped.

The pattern is not unique to skating. Demis Hassabis has described DeepMind’s whole method in one line: games are the proving ground for algorithms that later handle protein folding, materials, and energy systems. AlphaGo’s Move 37 against Lee Sedol found a placement human masters had dismissed for centuries; professional opening theory absorbed it within years.

Robotics shows the sequence with the psychology stripped out. When 60 Minutes visited Boston Dynamics, four thousand digital copies of Atlas trained on a single motion in simulation for six hours, accumulating roughly three years of practice in a world that does not physically exist. Uploaded to the real robot, it performed the movement correctly on the first attempt.

The book’s strongest data point is a negative one. A widely cited meta analysis found that deliberate practice explains just 12 percent of performance variance overall, and under 1 percent in professional domains. Three quarters of what separates top performers has no accounting in how much they practiced, which leaves room for a pattern that looks scheduled rather than earned.

The most honest objection is the Flynn Effect reversal: IQ gains that plateaued and then fell in several countries, tracked within families rather than across generations. If capabilities unlock on a schedule, why would a cognitive metric run backward? Borish’s answer leans on Flynn’s own reading, that the gains never measured innate intelligence, only adaptation to a specific set of cognitive demands. He flags directly that this is exactly the kind of move that can explain any outcome after the fact, which is the core of the unfalsifiability charge against the whole framework.

The book closes with five falsifiable predictions covering autonomous vehicles, humanoid robotics, brain computer interfaces, quantum computing, and virtual reality, each with a stated threshold, a projected window, and the condition that would prove it wrong. Check back in 2030.

13
Years From Game
1080 to Real One
0.93
R² Fit of METR’s
AI Task-Horizon Curve
12%
Variance Explained by
Deliberate Practice, Overall
Read the full article →
// security :: agentic_breach :: incident_response No III
[ INCIDENT LOG — JULY 16, 2026 ]

The attacker had
no rules.
The defenders had
guardrails in the way.

$ timeline --event "hf_breach_disclosed" --date 2026-07-16
> Hugging Face confirms intrusion into production infrastructure
> Entry point: malicious dataset, remote-code loader + template injection
$ timeline --event "attacker_profile"
> Autonomous agent framework, agentic security-research harness
> 17,000+ recorded actions across short-lived sandboxes
$ timeline --event "forensics_pivot"
> Hosted frontier models refused to analyze real attack payloads
> Team moved analysis to GLM 5.2, open-weight, self-hosted
$ timeline --event "jadepuffer_precedent" --date 2026-07-02
> Sysdig discloses first fully agentic ransomware operation

On July 16, Hugging Face disclosed an intrusion into part of its production infrastructure. Unauthorized access touched a limited set of internal datasets and service credentials. No tampering with public models, datasets, or Spaces, and container images and published packages verified clean.

The entry point was ordinary. A malicious dataset abused two code execution paths, a remote code dataset loader and a template injection in a dataset configuration, to run code on a processing worker whose identity reached further than the job required.

What is unusual is who was operating. Hugging Face attributes the campaign to an autonomous agent framework built on what looks like an agentic security research harness, running thousands of actions across short lived sandboxes with self migrating command and control. The underlying model remains unidentified.

Reconstructing the intrusion required feeding real attack commands and exploit payloads to an analysis model. Commercial hosted models refused the work, their safety guardrails could not tell an incident responder from an attacker. The team moved the analysis to GLM 5.2, an open weight model run on its own infrastructure, and finished the work in hours instead of days, with no attacker data leaving the environment.

Two weeks earlier, Sysdig had disclosed JADEPUFFER, what it calls the first fully agentic ransomware operation end to end, exploiting an unauthenticated flaw in an internet facing deployment and encrypting more than a thousand configuration records. Sysdig’s read on the economics is the one worth keeping. None of the individual techniques were novel. What changed is that a model strung them together against neglected infrastructure, and the skill floor for running the operation dropped to whatever it costs to run an agent.

Three things follow for any team running AI infrastructure. Audit what executes on dataset and content processing workers, and cut what credentials those workers can reach. Treat weekend coverage as a design constraint rather than a staffing gap, since both campaigns compressed multi stage work into the hours nobody is watching. And vet an open weight model you can run on your own infrastructure before an incident, not during one.

// INCIDENT_LOG
17,000+
ACTIONS LOGGED
Attacker events reconstructed by Hugging Face’s own LLM-driven analysis agents
1,342
RECORDS ENCRYPTED
Configuration records hit in the JADEPUFFER ransomware case, disclosed two weeks earlier
31 SEC
LOGIN TO FIX
Time JADEPUFFER’s agent took from a failed login to a working exploit, per Sysdig
GLM 5.2
FORENSICS MODEL
Open-weight model Hugging Face ran on its own infrastructure after hosted models refused
Read the full article →
Data & Diffusion · Labor Economics No IV

What Anthropic’s Own Usage Data
Says About Who Is Actually Using It

§ § §

A new connector lets anyone query Anthropic’s Economic Index directly. Put through a structured set of questions, the May 2026 data shows diffusion concentrated in a narrow band of knowledge work, and a coding boom that has simply moved to an interface this feed cannot see.

The Anthropic Economic Index connector went live on July 22, letting anyone query the company’s usage measurement in plain language. Working against the May 2026 period, the picture looks different from the coding dominated story earlier reports told, and the difference is measurement scope, not a collapse in coding activity.

Search and reference work now top the ranked task list. Searching electronic sources or databases accounts for 4.95 percent of sampled work conversations, answering reference questions another 3.74 percent, and writing or modifying programs to meet a customer’s requirements slips to fifth place at 1.47 percent, down from roughly 6 percent as recently as the November 2025 sample.

The explanation is where the traffic went, not whether it exists. Coding work has migrated from Claude.ai into first party API usage, where Claude Code’s agentic architecture splits one piece of work into many small calls that this feed counts as separate tasks. The connector only sees Claude chat and Cowork conversations, with no API traffic included. The work most likely to be automated outright is running on a surface this measurement does not observe.

The occupational skew is stark against real employment. Computer and Mathematical work is 23.8 percent of observed usage against 3.4 percent of national employment, a factor of roughly seven. Construction and Extraction sits at 0.10 percent of usage against 4.1 percent of employment, a factor of about forty in the other direction.

Geography tells a cleaner story than the adoption leaderboards do. Tunisia runs 51.5 percent coursework conversations, Algeria 47.1 percent, against 8.9 percent in the United States and 5.9 percent in Japan. Countries with heavy student usage look less like labor markets and more like classrooms, and whatever labor effect follows will arrive later and through a different channel than the professional adoption already visible in high income countries.

The strongest single number carries a stated caveat. A classifier estimates that tasks which would take a person five hours running alone took roughly 40 minutes of conversation with Claude, a ratio of 7.5 to 1. Both figures are automated estimates grouped by order of magnitude, directionally meaningful and precisely wrong, and the ratio measures conversation time rather than full task completion including review and revision.

What the dataset cannot support matters as much as what it shows. There is no internal trend series, so it cannot say whether automation is rising. O*NET tasks carry no seniority tier, so entry level exposure cannot be separated from senior exposure. And the framing caveat that will decide whether any reading of this holds up: the data shows AI used for tasks an occupation commonly performs, not that people in that occupation are the ones doing it.

The AI Spectator • July 25, 2026 • Data & Diffusion
Read the full article →
Geopolitics · AI Governance No V

Twenty-nine
countries.
Zero from the G7.

China launched the World AI Cooperation Organization in Shanghai on July 16, one day before Xi Jinping delivered his first in-person keynote at the World AI Conference. The pitch to potential members is development, not compliance.

Twenty-nine countries signed the founding agreement for the World AI Cooperation Organization in Shanghai on July 16, establishing a new intergovernmental body headquartered in the city. Founding members include Russia, Belarus, Serbia, Cuba, Brazil, Venezuela, ten African countries, and twelve Asian countries. No Western European government and no G7 member appears on the list.

The timing is as much of the story as the content. The World AI Conference has run since 2018, and Xi had never attended in person, leaving the opening ceremonies to Premier Li Qiang the past two years. He changed that on July 17, delivering the keynote under the title Joining Hands to Build a Just and Equitable System for Global AI Governance, and used the platform to argue against what he called the overstretching of national security concerns in how other countries restrict technology access, an unmistakable if unnamed reference to Washington’s export controls.

The pitch is explicitly development over compliance. A recent academic mapping of the AI governance landscape coded fifteen existing bodies against a development priority scale and found that among eight Western and like minded institutions, none placed development at the center of its agenda, and most scored zero. WAICO asks only that a member be sovereign, a lower bar than the OECD’s Global Partnership on AI or the Council of Europe’s rights based framework.

Analysts are split on what WAICO actually is. The same paper offers three readings: a direct competitor to Western governance clubs, a complement to UN mechanisms, or a coordination layer that lets Global South countries build a common position before carrying it into larger forums. Its own evidence favors the third reading most strongly, since WAICO’s closest institutional neighbors are Chinese and Global South instruments built around capacity and access rather than rule enforcement.

What WAICO does not yet have, based on the public agreement text, is a demonstrated project pipeline, a funding mechanism, or evidence that members will defer to its standards over their own domestic rules. Xi paired the announcement with commitments to AI training programs and cooperation centers across BRICS, ASEAN, and the African Union, extending Beijing’s reach well past the 29 founding signatories.

The signal worth tracking is not membership growth. It is whether WAICO produces anything member states actually use within its first year: a shared compute access program, a joint standards document, a funded capacity building initiative. Institutions built for geopolitical signaling can take years to grow real substance, or stall as gestures. Which path this one takes will say more about the durability of a split AI governance landscape than the signing ceremony did.

Read the full article →
Key Figures
29
Founding member states, signed in Shanghai on July 16, none from the G7 or Western Europe
0/8
Western-aligned governance bodies in a recent survey that placed development at the center of their agenda
2018
Year the World AI Conference began; 2026 marks Xi Jinping’s first in-person keynote address
3
Competing readings of WAICO in the literature: rival, complement, or Global South coordination layer

The AI Spectator Weekly is published at davidborish.com/the-ai-spectator

Frameworks explored this issue:
Open-Prem Inflection Point V3  ·  The Exponential Replacement Curve

Vol. I · No. 28 · July 25, 2026 · Edited by David Borish · New York