|
AI Builders Digest
Tuesday, November 10, 2026
|
|
An AI startup proved a 2008 quantum computing conjecture on day three of a run. The same week, a data company hit $1 billion in annual revenue selling training tasks that frontier models still fail at embarrassing rates. The story of AI progress right now is not one curve. It is two: what these systems can do at their ceiling is astonishing, and what they get wrong at their floor would make your intern wince.
|
|
---
|
|
01
|
An AI system just settled an 18-year-old quantum computing question
|
|
|
Physical Superintelligence, an AI science startup running a non-public Anthropic model, proved the Aaronson-Ambainis conjecture, a question open since 2008 about when quantum computers can be exponentially faster than classical ones. The system found a complete proof on day three of the run, then improved it twice more. A formal proof checker called Lean verified every step. Anthropic co-founder Jack Clark, writing in Import AI, called this a taste of the singularity: "mostly automated synthetic researchers extending the frontier of knowledge, spitting out results which need to then be understood by humans."
|
Why it matters: Yesterday's digest covered OpenAI benchmarking whether agents can read a lab notebook, and DeepMind proposing novel drug molecules. Add this, and you have a third data point in three days of AI crossing into territory that used to require a career's worth of human expertise. The question is no longer whether AI can do science. The question is how fast institutions, journals, and funding bodies can process results that arrive faster than any human committee can review them.
|
|
Source →
|
|
---
|
|
02
|
The company selling "hard AI tasks" hit $1B in revenue, and the demos show why
|
|
|
Handshake AI went from launch to $1 billion in gross annualized revenue in roughly a year by building training data and benchmarks for frontier labs. Latent Space's deep dive reveals what that actually means: tasks where models still fail in ways that matter. In a finance demo, a model correctly explains tax treatment when asked directly, then bungles applying that same rule to a specific line of code. In a biology demo, it misidentifies tissue samples in a brain tumor dataset that a trained human gets right on first look. In a coding demo, the evaluation catches not just what the agent returned, but which files it read and what it did to shell history.
|
Why it matters: Every lab selling you an AI product is buying data from companies like Handshake to patch exactly these failure modes. The finance demo is not an edge case. It is a description of how AI-assisted due diligence fails today: confident on the concept, wrong on the application. If your company is using an AI tool for financial analysis, the model almost certainly has a version of this bug that its vendor is actively trying to train away.
|
|
Source →
|
|
---
|
|
03
|
CodeRabbit says it ships 20 times more software using OpenAI's Codex
|
|
|
Code review platform CodeRabbit published a case study claiming it ships 20 times more software with the same engineering team by using OpenAI models to govern AI-generated code and Codex for the actual development work. The framing is about trust: not just generating code, but having a system that can evaluate whether AI-written code is safe to merge.
|
Why it matters: "20x" is a number that will end up in every AI ROI presentation your leadership team sees this quarter. The important word in CodeRabbit's claim is "govern." If the tool that reviews AI-generated code is itself AI, you have removed the human from the loop in both directions. That is worth asking about before the slide deck reaches your CFO.
|
|
Source →
|
|
---
|
|
04
|
Box CEO Aaron Levie on why agent swarms break the productivity ceiling
|
|
|
Aaron Levie argued that the shift from conversational AI to autonomous agents is a qualitative change, not a quantitative one. Chat-based AI produces output at roughly the speed of a back-and-forth conversation, which caps productivity gains somewhere around 20 to 50 percent. Agent swarms that fan out complex work in parallel have no equivalent ceiling.
|
Why it matters: This connects to Box's argument from yesterday that latency does not matter for long-running agents. Levie is building a consistent case that cloud-hosted, asynchronous, parallel agents are the actual enterprise product, and the sell is that there is no comparable on-premise alternative. Whether the 20x claim from CodeRabbit above is representative or cherry-picked is exactly the question that tests this thesis.
|
|
Source →
|
|
---
|
|
05
|
ChinaTalk's WarTalk: post-midterms defense fallout
|
|
|
This week's WarTalk covers the political aftermath of the midterm elections: the Iran conflict, the future of Pentagon leadership, and what Congress does to defense policy when a lame-duck dynamic sets in. No direct AI content, but useful context for anyone tracking how US defense and China policy might shift in the coming months.
|
|
Source →
|
|
Follow builders, not influencers. A daily digest of what matters in AI.
Read online ·
Archive
|