Nine items today, and most of them are arguments about attribution: who or what actually produced the result, and who ends up paying for it.
- Whose drafts. OpenAI claims a Navier-Stokes proof, and the mathematicians who fed their work into Codex want to know what trained it.
- A thousand engineers. Google is paying Accenture to place forward-deployed staff inside customers to make Gemini Enterprise work in production.
- $900 million and $800 million. Cognition raised at a $48 billion valuation with revenue and cash burn nearly matched.
- Friday. EU breach reporting starts September 11 with a 24-hour clock and no revenue threshold.
- The founders return. Dines at UiPath and Bhusri at Workday, on the argument that only a founder can authorize a rebuild.
- Ninety-five percent slop. Kevin Roose compressed a book from five years to one, then discarded most of what his AI reviewers handed back.
- Free is a line item. Lovable treats free inference as acquisition spend because AI gross margins fell toward 40%.
- 23.9 versus 82.2. An agent building an agent passes a quarter of the tests; the same model with an engineer passes most of them.
- 11x, and losing money. The return on ad spend was brand campaigns taking credit for buyers who had already decided.
The pattern worth pricing: the headline number usually belongs to someone other than the system reporting it.
Developments
AI Strategy & Transformation
September 8: OpenAI's math proof arrived with a question about whose drafts trained it
OpenAI said on September 8 that an unreleased internal model, run as roughly 10,000 parallel agents for 88 hours, produced a proof resolving the Navier-Stokes Millennium Prize problem. NYU's Tristan Buckmaster and Anthropic's Levent Alpöge had spent a year on the same unusual route, and Buckmaster says they had been putting their drafts into Codex sessions throughout. He published a statement questioning whether that usage informed OpenAI's result. OpenAI's Sébastien Bubeck called the allegation false and inflammatory, and the company said it cannot rule out that de-identified usage data helped improve its models.
So what
The claim to act on is not the proof, it is the sentence about de-identified usage data. Most growth companies push strategy documents, source code and customer records into a vendor's tools every day under terms that permit exactly this. Pull your three largest AI vendor agreements this week and read the training and telemetry clauses. If you cannot get a written exclusion, move the sensitive work to a tier that has one.
Hype check
OpenAI's proof is unpublished and has not been peer reviewed, while Buckmaster and Alpöge have released Lean-verified blow-up results for related fluid systems. Treat the mathematical claim as pending and the dispute as the reportable fact.
Confidence: Watch this — the announcement and the dispute are documented, but the proof's validity is not yet independently established.
Go deeper OpenAI · The Rundown AI
Go-to-Market
September 8: Google is paying Accenture to put 1,000 engineers inside its customers
Accenture and Google Cloud announced the Accenture Gemini Enterprise Business Group on September 8. Google Cloud will train up to 1,000 Accenture forward-deployed engineers who sit on site with customers to move Gemini Enterprise projects into production, drawing on Accenture's roughly 50,000 Google Cloud-certified staff. The companies cited a YouTube deployment during NFL Sunday Ticket demand peaks that raised customer sentiment 11% and cut average handling time 37%.
So what
This is the clearest signal yet that enterprise AI is being sold as a services-attached deal rather than a seat license. If your buyer now expects engineers on site, a self-serve motion and a two-day onboarding will start losing deals you used to win on product alone. Price and staff deployment explicitly, and decide whether you build that capacity or partner for it. The alternative is watching a systems integrator own the customer relationship you thought was yours.
Confidence: Established fact — announced by both companies with a named business unit and a headcount target.
Go deeper Accenture Newsroom · TechCrunch
Operations & Finance
September 8: Cognition raised at $48 billion with $900 million of revenue and $800 million of burn
Cognition, maker of the Devin coding agent, raised more than $2 billion at a $48 billion valuation on September 8, in a round led by Andreessen Horowitz with Accel, Founders Fund, General Catalyst and Avenir. Annualized revenue nearly doubled from $492 million in May to roughly $900 million. TechCrunch reports the company leases an Nvidia server cluster that could push total cash burn to about $800 million this year, and expects $4 billion to $5 billion of annualized revenue by the end of 2026.
So what
The growth is real and so is the cost structure sitting underneath it. A dollar of AI-native ARR is not the same asset as a dollar of SaaS ARR, because inference and multi-year compute leases land in cost of revenue rather than in a one-time build. When you benchmark your own multiple against rounds like this one, benchmark gross margin and compute commitments alongside growth rate. Ask your board to see burn per dollar of net-new ARR at the next meeting.
Confidence: Established fact for the round and valuation; the burn and forward revenue figures are press estimates rather than disclosed numbers.
Go deeper Cognition · TechCrunch
Regulation
September 11: EU breach reporting starts, and there is no revenue threshold
From September 11, 2026, any manufacturer placing hardware or software with digital elements on the EU market must report actively exploited vulnerabilities under the Cyber Resilience Act. An early warning is due within 24 hours of becoming aware, a fuller notification within 72 hours, and a final report within 14 days of a corrective measure. Severe incidents follow the same 24-hour and 72-hour steps. Reports go to the national CSIRT and ENISA through the CRA Single Reporting Platform, and products already on the market are covered.
So what
There is no size exemption, so a five-person company with one EU customer is inside the rule on Friday. Twenty-four hours makes this a runbook problem rather than a legal one. Name the person who files, write the decision test your team will use for "actively exploited," and register on the reporting platform before you need it. Do this even if you sell through a reseller, because the obligation follows the manufacturer.
Confidence: Established fact — the date and the reporting thresholds are set in the regulation and published by the European Commission.
Go deeper European Commission · National Law Review
Perspectives
People & Organization
The founders are coming back, and it is a statement about who can authorize a rebuild
SaaStr argues that founder returns at pre-AI B2B companies are becoming a pattern rather than a coincidence, pointing to Daniel Dines retaking UiPath in June 2024 and Aneel Bhusri returning to Workday in February 2026. The case is that AI forces a foundational product rebuild, and that professional CEOs hired to optimize a working model are the wrong operators for that job. The piece is candid that both returns have stabilized the business without yet reigniting growth.
So what
If your category is being re-architected, the scarce capability is institutional memory about why the product was built this way, not operational polish. Ask honestly whether your current leadership can authorize a rebuild that invalidates the roadmap they wrote and the targets they committed to. For investors, a founder's return is a signal that the board concluded incremental change had already failed. The useful follow-up question is whether the rebuild is actually funded and underway.
Confidence: Informed perspective — the founder returns are documented; the causal claim about AI is the author's argument.
Go deeper SaaStr · TLDR Founders
People & Organization
Kevin Roose ran his book past a council of AI reviewers and threw out 95% of what came back
Every interviewed five professional writers about where AI helped and where it did not. Kevin Roose delivered a 448-page book built on more than 150 interviews in roughly a year, and estimates it would have taken three to five years without AI handling research, transcription and fact-checking. He also refused to let a model generate a single sentence in it, and ran each chapter past a set of AI reviewer personas he calls his council of Claudes. Ninety-five percent of what they returned was, in his word, slop.
So what
The number that matters is the 95%, not the compression. The gain came from a practitioner who already knew what good looked like and could find the useful 5% quickly. Before buying an AI tool for a team, identify who holds that judgment and model what happens to throughput when they become the reviewing bottleneck. Roose also kept confidential material away from the lab whose company it concerned, which is the same discipline today's lead item argues for.
Confidence: Informed perspective — first-person practitioner accounts reported by Every, not measured results.
Go deeper Every · Every / Compound Writing
Resources
Product & Technology
An agent building an agent passed 23.9%. The same model with an engineer passed 82.2%
Sierra open-sourced Hyper-τ-Bench, which asks a developer agent to build a deployable customer-service agent from a simulated business's documents, transcripts, API and codebase, then tests the result on unseen customer conversations across airline, retail, telecom and banking. Claude Opus 5 at maximum reasoning, running in Claude Code, passed 23.9% of held-out tasks working alone. The same class of model paired with an engineer who had deep context passed 82.2% on the same tasks. The method is published on arXiv.
So what
This is the cleanest published measure yet of what the human contributes to an agent build, and it is most of the result. The gap is business context, not model capability. If you are budgeting an agent program, fund the person who knows the domain well enough to specify it, because that role is where the 58 points live. Take the benchmark's task structure and run it against one of your own workflows before you commit headcount.
Confidence: Established fact — a published benchmark with an open-source implementation and an accompanying paper.
Go deeper Sierra · arXiv paper
Go-to-Market
The account reporting 11x return on ad spend was losing money on every order
Search Engine Land walked through an account whose platform-reported return on ad spend was 11x and whose orders lost money once returns, VAT, cost of goods and fulfillment came out of what the platform counted as conversion value. Most of the 11x came from brand campaigns near 18x, largely shoppers who had already decided to buy, while real non-brand return sat closer to 3x. Correcting for margin alone created a second problem, because bidding purely for efficiency starved an apparel SKU of volume and left stock to be marked down below cost.
So what
Your ad platform reports revenue it did not earn and cannot see the costs that arrive after the order. Rebuild conversion values on contribution margin rather than gross revenue, and separate brand from non-brand before you judge either one. Then give each SKU a single job: profit, volume, or cash recovery before the season closes. The same failure mode is already showing up in AI dashboards that report throughput while the human review queue grows.
Confidence: Informed perspective — a practitioner case account rather than an audited disclosure.
Go deeper Search Engine Land · TLDR Marketing
