Frontier AI Architect

This episode's audio is AI-generated, based on an article written by David Amar.

Three US agencies, Anthropic, Google and OpenAI all describe the same thing in 2026: models copied through paid API access, not stolen files. This episode explains the three ways a model leaves, and how a startup should order its defences.

In this episode:

- The September 2026 joint advisory on industrial-scale distillation, and China's response

- What Anthropic, Google and OpenAI reported about campaigns against their own models

- Why each request looks normal: our per-account calculation

- The insider and customer cases: Linwei Ding at Google and the Mistral leak

- The case that most startups should not spend on this, and what to do at each stage

Read the full article with all sources: https://frontieraiarchitect.substack.com/p/nobody-needs-to-steal-your-weights?r=978zej&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

Frontier AI Architect is written by David Amar.

LinkedIn: https://www.linkedin.com/in/david-amar-575192441

Newsletter: https://frontieraiarchitect.substack.com

Contact: [email protected]

This episode's audio is AI-generated, based on an article written by David Amar.

Windsurf offered to pay in full and still lost nearly all of its direct Claude capacity with less than five days of notice. This episode explains why allocation, not price, is the real supplier risk in AI, and what the major providers actually promise in writing.

In this episode:

- How Windsurf lost its Claude capacity over an acquisition that never closed

- The four layers of dependence: model lab, serving channel, cloud and silicon

- What six provider pages promise: 45 days to six months before a model retires, no notice period before capacity changes

- Why a backup route needs live traffic: the ramp-rate calculation

- The case for committing everything to one supplier, and why the largest buyers do both

Read the full article with all sources: https://frontieraiarchitect.substack.com/p/your-ai-supplier-can-drop-you-in?r=978zej&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

Frontier AI Architect is written by David Amar.

LinkedIn: https://www.linkedin.com/in/david-amar-575192441

Newsletter: https://frontieraiarchitect.substack.com

Contact: [email protected]

This episode's audio is AI-generated, based on an article written by David Amar.

Agents can already pay on cards and on stablecoins, but the measured volume is small and no rule yet says who bears the loss when an agent errs. This episode walks through the three layers of an agent payment and where each one breaks.

In this episode:

- Why major US banks restricted third-party AI agents in October 2026, and what Walmart did instead

- What the Visa and Artemis data shows: $15.0 million of adjusted x402 volume in about eleven months

- The three layers: who the agent is, what the user authorised, and how the money moves

- Why the protocols produce evidence but leave the loss unassigned

- The case that machine payments are an early curve, not a flat one

Read the full article with all sources: https://frontieraiarchitect.substack.com/p/agents-can-pay-who-pays-when-they?r=978zej&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

Frontier AI Architect is written by David Amar.

LinkedIn: https://www.linkedin.com/in/david-amar-575192441

Newsletter: https://frontieraiarchitect.substack.com

Contact: [email protected]

This episode's audio is AI-generated, based on an article written by David Amar.

OpenAI and Anthropic both got the reasoning setting wrong for their own flagship products, in opposite directions. This episode explains what extra thinking buys, where it stops paying, and how to set it per task.

In this episode:

- GPT-5's router on launch day and the Claude Code effort default that Anthropic reversed

- Why accuracy grows with the logarithm of thinking tokens while cost grows in a straight line

- Why the gain sits in the middle band of task difficulty

- Our calculation from October 2026 list prices: effort moves one request's cost 23 times

- The case that adaptive thinking makes the dial obsolete

Read the full article with all sources: https://frontieraiarchitect.substack.com/p/how-long-should-your-model-think?r=978zej&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

Frontier AI Architect is written by David Amar.

LinkedIn: https://www.linkedin.com/in/david-amar-575192441

Newsletter: https://frontieraiarchitect.substack.com

Contact: [email protected]

This episode's audio is AI-generated, based on an article written by David Amar.

An appeals court said the shopper, not Perplexity's agent, is the one accessing Amazon, and Amazon is now fighting on through its own terms of service. This episode explains the four ways an agent gets into someone else's app, and who holds the off switch in each.

In this episode:

- What the Ninth Circuit decided in Amazon v. Perplexity, and what it left open

- Amazon's amended complaint: a contract claim and a cloud-run version of Comet

- The four routes into an app: declared intents, tool protocols, commerce protocols and screen-driving

- How JPMorgan and Walmart answered the agent question in October

- The case that screen-driving agents make the platforms irrelevant

Read the full article with all sources: https://frontieraiarchitect.substack.com/p/amazon-v-perplexity-who-is-the-user?r=978zej&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

Frontier AI Architect is written by David Amar.

LinkedIn: https://www.linkedin.com/in/david-amar-575192441

Newsletter: https://frontieraiarchitect.substack.com

Contact: [email protected]

This episode's audio is AI-generated, based on an article written by David Amar.

The first AI sales agents automated the one part of outbound that was already failing: sending more messages. This episode explains why volume became a liability, how buyers now ask a model first, and what a sales agent is good for.

In this episode:

- What happened at 11x, from the public reporting

- How mailbox providers' spam thresholds punish volume

- What the research on generative engine optimization shows, and what it does not

- Where a sales agent helps and where it hurts

- How the approach changes with deal size, from four-figure to enterprise contracts

Read the full article with all sources: https://frontieraiarchitect.substack.com/p/why-ai-sdrs-didnt-fix-outbound?r=978zej&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

Frontier AI Architect is written by David Amar.

LinkedIn: https://www.linkedin.com/in/david-amar-575192441

Newsletter: https://frontieraiarchitect.substack.com

Contact: [email protected]

This episode's audio is AI-generated, based on an article written by David Amar.

A two-person telehealth company reported $401 million in first-year sales, along with an FDA warning letter and several lawsuits it disputes. This episode looks at what that case proves about very small AI-native companies, and what it does not.

In this episode:

- What the public record says about Medvi, and what the company disputes

- Why the forecast of a one-person billion-dollar company came true in an unexpected category

- Three layers of a very small company: agent production, rented licensed functions, accountability

- Why responsibility for affiliates, vendors and chatbots stays with the company

- The counter-argument that the model simply worked

Read the full article with all sources: https://frontieraiarchitect.substack.com/p/what-a-401m-two-person-company-proves?r=978zej&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

Frontier AI Architect is written by David Amar.

LinkedIn: https://www.linkedin.com/in/david-amar-575192441

Newsletter: https://frontieraiarchitect.substack.com

Contact: [email protected]

This episode's audio is AI-generated, based on an article written by David Amar.

Stability AI's founder was still the majority shareholder when he left a company whose compute bill was forecast at nine times its revenue. This episode explains how compute spend turns into equity and what each way of paying for it costs. It describes how these deals are structured and is not financing advice.

In this episode:

- The Stability AI forecast and the reported terms of its rescue

- The exchange rate between compute spend and points of the company

- Five ways to pay: equity cash, credits, commitments, GPU-backed debt and vendor capital

- How the right choice changes from seed to growth stage

- The evidence that AI founders are being diluted less, not more

Read the full article with all sources: [LINK]

Frontier AI Architect is written by David Amar.

LinkedIn: https://www.linkedin.com/in/david-amar-575192441

Newsletter: https://frontieraiarchitect.substack.com

Contact: [email protected]

This episode's audio is AI-generated, based on an article written by David Amar.

Days before the EU AI Act's high-risk rules were due to apply, the EU pushed them back by up to sixteen months. This episode explains what moved, what is already in force, and which engineering work is safe to build now. It is engineering guidance, not legal advice.

In this episode:

- What the Digital Omnibus on AI changed, and the new dates

- What already applies: prohibitions, general-purpose model duties and Article 50 transparency

- Provider or deployer: who owes what, and how roles can flip

- Five artefacts that no amendment has changed

- The case for waiting until the standards are finished

Read the full article with all sources: https://frontieraiarchitect.substack.com/p/the-eu-ai-act-moved-your-work-didnt?r=978zej&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

Frontier AI Architect is written by David Amar.

LinkedIn: https://www.linkedin.com/in/david-amar-575192441

Newsletter: https://frontieraiarchitect.substack.com

Contact: [email protected]

For a fixed level of quality, token prices fall about tenfold a year, and AI bills keep rising. This episode explains why, what running your own GPUs really costs, and how to decide per workload.

In this episode:

- Why Cursor and Anthropic both reworked their flat-price plans in 2025

- Tokens per task and model tier: the two numbers that drive the bill

- The six costs of ownership, and a worked break-even example

- Two software levers to pull before buying hardware: the serving stack and routing

- Two opposite objections: wait for prices to fall, or own everything

Read the full article with all sources: https://substack.com/@frontieraiarchitect/note/p-221644919?r=978zej

Frontier AI Architect is written by David Amar.

LinkedIn: https://www.linkedin.com/in/david-amar-575192441

Newsletter: https://frontieraiarchitect.substack.com

Contact: [email protected]

This episode's audio is AI-generated, based on an article written by David Amar.