Ten items today, and the through-line is decomposition: the frontier model is coming apart into pieces you can own.
- Salesforce trained its own. Koa is a CRM reasoning model post-trained on Nvidia's open Nemotron 3 Super.
- A model that will not write. TypeSafe's Jev returns typed decisions at $42 per billion input tokens, output free.
- Congress has a bill. OpenAI backed a House mandate for outside evaluators, and a paid audit market opened.
- Materials science, 1,300 GPUs. Periodic post-trained an open trillion-parameter model and beat Astra on its benchmark.
- Two trillion, adjusted. Anthropic targets a Nasdaq listing while its profit excludes training costs.
- The middle disappears. Two investors said the same day that AI hollows out mid-sized software.
- Slop grenades. Shopify's chief executive named the cost of cheap output: the reviewer pays it.
- 1,393 subagents, $19,300. Nous refactored a million lines, and human review caught what the tests missed.
- Audit your own folders. Every published the prompt that makes an agent grade your workspace.
- Fight, influence, generate. A framework for reallocating paid search as AI Overviews absorb informational intent.
The pattern worth planning against: the general model is becoming a component, and margin moves to whoever owns the domain.
Developments
AI Strategy & Transformation
September 15: Salesforce trained its own reasoning model rather than renting one
At Dreamforce on September 15, Salesforce and Nvidia introduced Koa, a reasoning model for CRM work built by post-training Nvidia's open Nemotron 3 Super. The training corpus is entirely synthetic, simulating personas such as irate support callers and negotiating sales reps across more than a dozen industries, so no customer data is involved. Salesforce reports that Koa matches or exceeds leading models on CRM actions with three times fewer errors, and hosts the model itself so requests stay inside its own systems. It ships alongside AIforce, an interface layer that replaces the traditional Salesforce user interface.
So what
The build-versus-rent question now has a third answer: post-train an open base model on the domain you already own. Salesforce is not becoming a frontier lab; it is buying an open base and adding the one input a lab cannot replicate, which is three decades of process knowledge. If your company holds proprietary workflow data, price that option before you sign another multi-year frontier contract. A useful question for your next vendor review is which part of a supplier's intelligence it owns and which part it resells.
Hype check
Every Koa performance figure is Salesforce-originated, measured on Salesforce's own CRM Benchmark with Salesforce running the comparison models. Nothing has been independently replicated, the model is in pilot, and general availability is not until winter 2026.
Confidence: Established fact — the launch, architecture and training method are documented, while the performance claims remain vendor-run.
Go deeper Salesforce announcement · Salesforce AI Research CRM Benchmark
Product & TechnologyDeveloping
September 15: TypeSafe shipped a frontier model that cannot write a sentence
Diogo Almeida, a ChatGPT co-creator who spent two years in stealth, brought TypeSafe out with Jev, which the company calls a System One Model. Jev takes a structured question and returns a typed answer carrying a calibrated probability, trained by a method TypeSafe names reinforcement learning from calibrated decisions. Input costs $0.042 per million tokens with output free, and end-to-end latency runs 70 to 500 milliseconds. The trade is explicit: Jev gives up free-form text generation and chooses only among options defined in advance.
So what
Most production AI spend is not conversation. It is classification, routing, scoring and screening, and that work has been paying language-model prices. If you run high-volume judgment inside your product, price the decision path separately from the generation path this quarter and measure what share of your token bill is actually a lookup. The constraint to test first is cardinality, because anything requiring an open-ended answer still needs a language model behind it.
Hype check
The comparison multiples circulating this week are inconsistent because they are drawn from different tables in TypeSafe's own post, which concedes its workflow evaluations were built by its own model capabilities team and that some bias could exist. The claim that Jev cannot hallucinate is asserted from schema guarantees rather than measured, and the company states plainly that it cannot prove its pricing is unsubsidized.
Confidence: Established fact — the launch, the pricing and the structural limits are documented by the vendor, and none of the performance multiples are independently replicated.
Go deeper TypeSafe announcement · Superhuman AI
Legal & RegulatoryDeveloping
September 15: OpenAI backed a bill mandating outside evaluators, and a paid audit market opened underneath it
At a Capitol Hill roundtable on September 15, OpenAI global affairs chief Chris Lehane endorsed the third-party assessor provision of the bipartisan House FRONTIER Act, sponsored by Representatives Jay Obernolte and Lori Trahan. The provision would require an independent organization to certify that a lab's safety protocols mitigate catastrophic risk to acceptable levels. OpenAI has not endorsed the wider bill. The same day, AIUC raised a $40 million Series A led by Ribbit Capital for AIUC-1, an agent-safety certification modeled on SOC 2 and built with roughly 250 security and risk leaders, with Cursor, Lovable, Harvey and ElevenLabs as customers.
So what
Yesterday's voluntary evaluator framework became a legislative proposal and a purchasable certificate in the same week. For anyone selling AI into regulated buyers, an agent-safety attestation is heading toward the slot SOC 2 occupies in the security questionnaire. Ask your three largest enterprise customers now whether they intend to require one in 2027, and put the audit in next year's budget before procurement asks for it.
Confidence: Established fact — the endorsement, the bill sponsors and the AIUC raise are each reported by named outlets, though the bill itself has not moved.
Go deeper CBS News on the FRONTIER Act · TechCrunch on AIUC
Product & Technology
September 15: Periodic Labs post-trained an open model on its own lab data and beat the frontier on cost
Periodic Labs released Neon, built by post-training Kimi K2.6, an open-weight trillion-parameter model, on proprietary experimental data generated in its own high-throughput materials labs. On FrontierXRD, Periodic's internal X-ray diffraction benchmark, success rose from 2.7% for the base model to 55.3%. Periodic reports that Neon outperforms GPT-6 Astra and Claude Fable 5.1 on scientific analysis at lower cost per analysis, and that its final training run used 1,300 H200s against the 100,000-plus Blackwell GPUs reported for Astra.
So what
This is the Salesforce move in a different industry, and together the two define the play: an open base model plus a data asset nobody else can buy. The capital requirement is a training run rather than a research lab, which puts it inside reach for a well-funded growth company. If your business generates proprietary outcome data, the board question is whether that data is worth more as a product feature or as post-training fuel. The honest caveat is that scaffolding matters as much as weights, so budget for the harness too.
Confidence: Established fact — Periodic documents its method unusually openly, though FrontierXRD is its own 134-sample benchmark scored by a model-judge ensemble, and every competitor was run inside Periodic's own harness.
Go deeper Periodic Labs on Neon · Periodic Labs on building labs that learn
Operations & FinanceNew Angle
September 15: OpenAI drew a $1.2 trillion approach while Anthropic's profit came with three exclusions
The Financial Times reported OpenAI in early talks for a pre-IPO private round above $1.2 trillion, with investors approaching the company rather than the reverse, and noted that the talks have not progressed far and the figure may not hold. Anthropic is separately reported to be targeting a Nasdaq listing at roughly $2 trillion, against a last private mark of $965 billion set in May. Its reported second consecutive profitable quarter is an adjusted operating profit that excludes model training costs, revenue sharing with partners including Amazon, and stock-based compensation.
So what
Private AI marks are being set by inbound capital and adjusted metrics rather than by anything a public-market investor could underwrite. If you are raising against AI comparables, expect diligence on which costs your own margin excludes, because the adjusted-profit conversation has now moved from private decks into the press. The question to put to your finance chief this month: would our gross margin survive if the compute we amortize were expensed in the quarter it was consumed?
Hype check
An operating profit that excludes training, revenue share and stock compensation is not profitability in any conventional sense, and short seller Jim Chanos publicly mocked the figure as cost-adjusted EBITDA. The $2 trillion is a targeted listing valuation rather than a priced round, and reported raise ambitions vary widely across outlets.
Confidence: Informed perspective — the reporting is credible, but every number here is a target, a talk or an adjusted measure.
Go deeper Bloomberg on the OpenAI round · MarketWatch on the profitability caveat
Perspectives
AI Strategy & Transformation
Two investors argued on the same day that AI removes the middle of the software market
Sequoia partner Mike Vernal published “Moats and the Barbell-ification of Software” on September 15, arguing that as the cost of writing and copying code approaches zero, moats built on code complexity, switching costs, migrations and custom integrations weaken. He expects the market to polarize into a barbell: a few large platforms winning enterprise on capital, brand and suite breadth, and a long tail of focused, AI-built micro-tools, with mid-tier point solutions disappearing between them. Dan Hockenmaier published “Momentum is Not a Moat” the same day, separating execution speed from durable defensibility built on data effects, scale economies or unique process.
So what
If your company is a mid-sized point solution, this is the thesis your next investor and your next acquirer are both reading this week. The positions both authors leave intact are platform breadth at the top and proprietary data or process at any size, which is precisely the asset Salesforce and Periodic monetized in the items above. Write down, in one sentence, the thing a competent team with agents could not rebuild in a quarter. If that sentence does not come easily, you have found this quarter's strategy work.
Confidence: Informed perspective — two independent investors making adjacent arguments on the same day is a signal about sentiment, not evidence about outcomes.
Go deeper Mike Vernal on barbell-ification · Dan Hockenmaier on momentum and moats
People & Organization
Shopify named the cost of cheap output, and it lands on the reviewer
On The Knowledge Project on September 15, Shopify chief executive Tobi Lütke said laziness has inverted at his company: it used to look like too little output and now it looks like too much. Someone lets an agent open a pull request, approves it unread, and the colleague reviewing it burns hours the sender saved. Shopify calls that an AI slop grenade. His paired argument is accountability, since a machine cannot be held responsible and the person who sent the work still can. Every's head of evals Mike Taylor made an adjacent case the same week in “13 Beliefs About AI Writing.”
So what
Generation cost collapsed and review cost did not, so review is now the binding constraint on knowledge-work throughput. The fix is a norm rather than a tool: whoever sends AI output owns its correctness, and unreviewed output is not a deliverable. Put that sentence in writing before your next all-hands, because the alternative is that your most senior people quietly become an editing pool. Measure it by asking managers where their week actually went.
Confidence: Informed perspective — Lütke's account of Shopify's internal norms is credible and unaudited, and the term itself may have originated elsewhere at the company.
Go deeper The Knowledge Project episode · Mike Taylor on AI writing
Case Studies
Product & Technology
September 2: 1,393 subagents refactored a million lines for $19,300, and human review caught what the tests missed
Nous Research ran 1,393 Hermes subagents, peaking at 218 concurrent and using Claude Fable 5.1 for inference, against its own million-line Python codebase over roughly 19 active hours. Non-test Python fell 34.4%, from 1,063,826 lines to 698,363, functions longer than 300 lines dropped from 192 to two, and average tokens per code lookup fell from 2,218 to 993. The main run cost about $19,300, or roughly $25,000 including follow-up sessions, against a manual estimate of $150,000 to $1.8 million. Two rounds of human review caught real regressions the existing tests had missed, including an automated rewrite that changed exception handling at about 65 sites.
So what
The transferable finding is not the cost ratio. It is that the human review layer found defects the automated test suite did not, which is the control that made the economics real rather than theoretical. Before authorizing a large agent run against production code, fund the review explicitly as a budget line rather than as an afterthought. Note also that the run required post-merge fixes, so treat the throughput figure as a ceiling and not a plan.
Confidence: Established fact — Nous published the run, the costs and its own failure modes, and the work dates from September 2 rather than this week.
Go deeper Nous Research write-up · The Neuron
Resources
Learning
Every published the prompt that makes an agent grade your own AI workspace
Every launched a recurring internal series in which one employee walks the team through their AI setup, with Cora general manager Kieran Klaassen as the first subject. His layout separates a context store from a dispatcher, organizes memory by time scale, and keeps an always-on machine so agents continue working after his laptop closes. The more portable artifact is the audit itself. Katie Parrott had an agent map her own workspace against coherence, feasibility, scope and adversarial failure modes, and it surfaced 25 top-level folders, more than a thousand raw session files, and three versions of one project across two machines.
So what
Most executives now have an AI workspace that accreted rather than got designed, and the cost shows up as agents loading the wrong context and producing confidently wrong work. Run the audit against your own folders this week and ask for the three most consequential problems with the smallest fix for each. The rule worth adopting from Klaassen's setup is that captured context requires human review before it becomes shared guidance.
Confidence: Informed perspective — one team's architecture published as a practice, and the underlying folder essay dates from April rather than this week.
Go deeper Every on workspace audits · The folder is the agent
Go-to-Market
September 15: A budget framework for paid search once the AI answer does the educating
Writing in Search Engine Land, bigmomo co-founder Ana Kostic argues that a single bidding strategy across all queries stops working once AI Overviews absorb informational intent, and proposes sorting terms into three buckets. Fight covers bottom-funnel commercial intent, where you pay for absolute top placement. Influence covers informational terms where the AI answer educates the buyer, so the objective becomes citation share and coordination with organic rather than outbidding your own listing. Generate Demand covers zero-click terms, where budget moves to YouTube, Demand Gen and first-party content ahead of the search.
So what
Your paid search budget is probably still allocated on a pre-AI-Overviews intent model, which means you are paying top-of-page rates for queries that now resolve inside the answer. Pull an intent split of your twenty highest-spend terms this month and move the informational share. One corroborating data point: a July SE Ranking study of AI Mode commercial queries found ad likelihood rose with cost per click, at roughly 54% above $10 and 24% below $2, so paid placement survives where the terms are expensive.
Confidence: Informed perspective — the framework is one agency's practice, and its widely quoted 83.3% informational figure comes from a single client account rather than a study.
Go deeper Search Engine Land · TLDR Marketing
