AI Echo - Inaugural Edition - June 10-17, 2026

Over a long enough timeline, we all become forward deployed engineers

Welcome to the first issue of AI Echo.

I’m Kevin Selhi, a philosopher and technologist. I’ve spent my career helping leaders spread new technologies at inflection points.

Insofar as I am a philosopher, I walk the path of Aristotle over Plato or Socrates. Socrates’s student was Plato. Plato’s student was Aristotle. Aristotle’s student was... Alexander the Great. I hope reading this newsletter will help you be great.

Although Aristotle is better known for Nichomachean Ethics, I’m a big fan of his practical ethics. In particular, its focus on “dirty hands problems.” I launch this newsletter in that spirit. Within it I’ll share a bit about how I’ve gotten, and continue to get, my hands dirty with technology.

My first job out of college was in media analysis. I worked in the New York office of a European consultancy whose clients were mostly German and Swiss financial services firms. I’d just come home from Oxford, where I mostly read philosophy and researched artificial intelligence. For the job, I had to get good at scanning large stacks of international and regional broadsheets and business wires.

I noticed that some papers, like The New York Times, cover important events with large stories co-authored by multiple journalists, but emphasize prose over signal to noise ratio. The editor’s hand is strong. Others, like The Financial Times, tend to cover major events with fewer words, across multiple stories, including different by-lines from reporters spread around the globe. Occasionally, none of the major dailies or weeklies agreed on the most important stories of the moment or the most significant tidbits from each.

I wished for a single digest that could cover all the most important and novel stories, from multiple angles, while appreciating that an obscure item buried in a nether section of some regional conference proceedings could end up the most important story of the decade. This newsletter is my attempt to fulfill that wish.

Recently, I built two AI newsletter research agents:

• Alpha was written with Anthropic’s Claude Code, and nests Codex 5.5, Opus 4.8, Sonnet 4.6, and Haiku 4.5
• Beta was written with OpenAI’s Codex, and nests all those same models for different tasks

Both Alpha and Beta conduct research, extract snippets from sources, grade them, compile stories, fact-check everything, draft a newsletter, and then format it for easy reading on mobile.

Beta does it for about 10% the price of Alpha, on a build and per draft basis. This points to a recurrent theme which Beta correctly surfaced, in a post from OpenRouter, who show that fusion models perform better than frontier models alone, at a fraction of the price.

I had originally been holding this project back for Anthropic’s Fable 5, the first Mythos-class model which came out briefly last week. Within days of its release, the lab was forced to retract its model under pressure from The White House. At the time, Fable 5 topped all benchmarks and leaderboards. Before I could begin working on Saturday, the model had been withdrawn from public use. More peculiarly, non-American employees at Anthropic were specifically banned from using it. This is clearly the most important story in AI from the past week. Beta, the cheaper Codex agent, missed it on every pass. Below is included Claude Code’s rather nuanced output. There’s recently been more “behind-the-scenes” reporting from Axios, which suggests a clash of cultures or personalities, where Republican insiders don’t understand the technology and blame the scientists for making them feel stupid.

Claude Code captured more than 50% of the coding agent market within 8 months of launch. Even with premium pricing and forcing more processes into their toll-based API, they dominate the market. Fable 5 was more than 3x the price of GPT5.5 high on a cost per task basis according to Artificial Analysis, and Opus 4.8 is almost 2x more expensive than the best OpenAI model. Neither Alpha nor Beta pointed out this fact for Echo readers, although the more expensive Alpha agent from Claude Code surfaced an interesting story about tokenomics, quoted below.

These are not just idle questions. Another major story of the week, besides the SpaceX IPO, was that company’s $60 billion acquisition of Anysphere, makers of Cursor. Their product was the previous darling of the coding agent space, and helped them make history as the fastest software company to hit $1 billion in revenue, 24 months after initial launch. “Cursor’s developer market share dropped from 41% to 26% in the last twelve months. And revenue doubled to $4B ARR in the same period.” per Aadith Sheth on X.com. Obviously the space is still hot. A few days ago, Cursor released their own model, Composer 2.5. It was built atop Moonshot’s Kimi K2.5 and debuted third on Artificial Analysis’s Coding Agent Index at 10% cost per task versus Opus 4.7 max. This is especially significant, not just because The White House can pull your base model on a whim. Although Anysphere is great at generating revenue, they’re even better at burning runway and margin through massive token spending. According to SaaStr, “one investment firm’s analysis suggested Cursor was paying $650 million a year to Anthropic against $500M revenue, i.e. a negative ~30% gross margin.” Now that they are in-house at SpaceX, we’ll see whether that means Grok gets better or Composer gets worse.

Although that should be all from me for now, I can’t help but make a couple more observations that weren’t picked up by either Alpha or Beta, but should be relevant to the most entrepreneurial Echo readers. Future issues will probably have less commentary from me and more from Alpha, Beta, and any other siblings that I birth for them in the coming weeks and months. But i can’t help but think about Groq (an incredible company at the bare metal layer that sold to NVIDIA in December) after mentioning Grok. Anysphere was founded by four former MIT students in 2022. They exited four years later for $60 billion. Groq was founded in 2016 by the guy who created the TPU at Google. Took twice as long to exit for one third the price. All cash though. What’s bizarre to me is the difference in valuation given the difference in margins. The Anysphere founders focused on the ultra low margin application layer. Windsurf margins were “very negative” prior to their awkward acquisition / acquihire by Cognition / Google and the same was true at Cursor until Composer.

I get that the application layer is hot and coding applications in particular are important for foundation model companies to master to speed their own flywheel. We don’t need to look any further than Edwin Chen and Surge AI to understand the significance of the particular application, if manifest on a different layer. It also makes sense of what has been going on under Alexandr Wang’s leadership, since the Scale AI acquihire, Meta engineers have become data labelers. But as regulators force foundation labs to slow down while consumers wake up to the capabilities of the available models and their adoption continues to explode, it stands to reason that LPUs and other hardware built for inference like the Groq chips will become even more valuable than GPUs.

My intuitions might be wrong here, but I get the sense that Groq will be a much better deal for NVIDIA than Anysphere ends up being for Grok/SpaceX.

- Kevin

The top line

Washington restricted Anthropic’s two newest frontier models — and turned vendor concentration into the week’s central deal risk. The Trump administration imposed emergency export controls on Claude Fable 5 and Mythos 5, forcing Anthropic to disable them within days of launch over national-security concerns (misuse, jailbreakability, export rules, foreign-national access); the government separately asked Anthropic to bar non-US users from its frontier Claude models, and Anthropic dispatched senior engineers to D.C. to negotiate ( NYT, Jun 13, 2026; WSJ, Jun 2026; Fortune, Jun 15, 2026; Al Jazeera, Jun 14, 2026). The restriction hits the latest tier, not the whole Claude line; reporting puts cut-off users as far as UK adviser Rishi Sunak and Indian enterprises, with one foreign-language account claiming the White House gave ~90 minutes to disable Fable 5 after Amazon-raised concerns, and The Verge/Semafor citing possible China-linked access to Mythos ( Yahoo News, Jun 2026; The Verge, Jun 2026; TechCrunch, Jun 13, 2026). Security leaders from Adobe, Zoom and Sophos are pressing for reversal — arguing the cyber capabilities already exist in rival models and the ban disadvantages defenders — while the UK, Canada (PM Carney), and France (”the AI war”) reframe it as proof of US-dependence risk ( Axios, Jun 15, 2026; The Times, Jun 2026; Le Monde, Jun 14, 2026; AP, Jun 2026).

• Why it matters: this is a live regulatory dispute, not settled market structure. The honest scope is ”the newest tier may be geofenced,” not ”your existing Claude deployment shuts off tomorrow.”

• The buyer ask is concrete: for any global/non-US workload, confirm whether your users sit inside the permitted geography, and whether your contract treats a government-ordered shutdown as force majeure or an SLA breach.

• The real conversation is ”is your architecture substitutable,” not ”is Anthropic safe” — push for portability clauses, contractual exit + data-egress rights, a vendor-neutral eval harness, and a pre-built governed failover that preserves audit logging, data residency, and your DPA. “Model-agnostic” on its own is not an answer a regulated buyer accepts; a tested switchover path is.

• Expect sovereign-AI mandates (EU, UK, Canada, India) to gain momentum — a tailwind for portability planning, not a reason to abandon a vendor mid-pilot.

• Flags: much of the detail (90-minute deadline, China-link, negotiation status, “without warning”) is reported/unconfirmed; duration and appeal path are not public.

Salesforce is buying AI customer-service agent Fin for ~$3.6B — its biggest deal since Slack. The acquisition (Fin was formerly part of Intercom) strengthens Agentforce; the stock rose as investors read it as a defense against AI eroding the SaaS model ( CNBC, Jun 15, 2026; TechCrunch, Jun 15, 2026; Barron’s, Jun 2026).

• Why it matters: the size of the cheque is the signal — Salesforce will buy agentic CX rather than build it, which means it now sees autonomous customer-service as a contested platform category, not a feature.

• For any Agentforce-adjacent deal: a buyer must weigh Salesforce’s newly-reinforced native stack against a best-of-breed integration — but a $3.6B deal also implies the incumbent lacked a credible in-house answer until now, so question day-one maturity.

• Treat it as a procurement checklist, not a capability guarantee: integration cleanliness, CRM write-permissions, are agent actions logged and reversible, data residency, whether pricing shifts to per-conversation/per-action, and whether outputs are covered by existing support + indemnity.

• Independent agent vendors now face pressure to differentiate sharply or become acquisition targets — flag that to customers betting on a smaller vendor’s independence.

• Flag: early issues carried this as single-source/unconfirmed (Investor’s Business Daily); now corroborated by CNBC + TechCrunch, but deal terms still warrant a primary Salesforce filing.

KPMG published, then pulled, an AI-generated report full of apparent hallucinations — after ChatGPT and Gemini had already cited it. The fabricated content propagated into both AI systems and a national newspaper and trade press before retraction ( TechCrunch, Jun 13, 2026; PCMag, Jun 2026).

• Why it matters: the cleanest governance cautionary tale of the week — and it lands precisely because the victim is an assurance firm. The failure was process, not model: no human-review gate strong enough to catch fabrication before external publish.

Use it to make human-in-the-loop review, citation-grounding, and provenance tracking non-optional in any content-generation pilot — the same controls procurement should be asking you to demonstrate.

Trust & governance

Malicious IDE plugins steal AI keys. Researchers found 15 malicious JetBrains marketplace plugins disguised as AI coding tools that exfiltrate OpenAI, DeepSeek and SiliconFlow API keys, plus Chrome extensions capturing chatbot conversations ( The Hacker News, Jun 2026). The attack surface is the developer workstation — high-privilege and under-monitored at exactly the companies rushing to adopt coding agents. Raise plugin governance in any IDE-integration POC.

G7 turns to AI sovereignty. G7 discussions on frontier risk, infrastructure sovereignty and U.S. dominance drew Sam Altman, Demis Hassabis and other executives, with European nations pushing for checks on American AI ( CNBC, Jun 17, 2026). The same regulatory current running through the Anthropic shutdown — and a signal of divergence that will complicate multinational deployments.

Huang calls for “new social norms.” Nvidia CEO Jensen Huang said society needs new norms around AI alongside sensible regulation and more data-center energy ( AP, Jun 2026) — useful framing for executive stakeholders, even from an interested party.

From the field

The tokenomics bill is coming due. Wired reports two enterprises — a software maker and an ecommerce company — are facing token consumption they call “pretty crazy,” sometimes 8x baseline estimates, forcing leaders to rethink ROI ( Wired, Jun 2026). As agentic workflows multiply calls, per-token pricing produces cost curves procurement never modeled. This belongs in every POC scoping conversation — set a token budget and an alerting threshold before, not after, the pilot scales.

OpenRouter’s Fusion API blends multiple models behind one endpoint, claiming near-frontier quality at ~half the cost. Reported 64.7% on a Perplexity benchmark vs Fable 5’s 65.3%; 194 points on HN ( OpenRouter; OpenRouter Fusion).

• Why it matters: a demoable cost-quality and resilience-against-single-provider-disruption pattern — timely against the Anthropic restriction. Treat as a POC candidate, not a production black box: routing raises which-model-saw-which-data logging questions. Flag: vendor self-reported benchmark.

MiniMax launched M3 — open-source, with coding reportedly approaching Anthropic’s Opus 4.7. Intensifies the open-weights coding race alongside Qwen and DeepSeek ( The Information, Jun 2026; llm-stats).

• Why it matters: a credible option for self-hosted dev tooling — but parity claims need independent validation, and Chinese-origin models carry their own procurement/data-governance scrutiny. Flag: reported, single-source, vendor benchmark.

NVIDIA Nemotron 3 Ultra (~550B, fully permissive) and Arcee AI’s Trinity (400B) widen the open-model field. Nemotron is pitched as a most-capable open model; Trinity is a 30-person US startup’s 400B model claiming to beat Llama — though Trinity dates to late January, not this week ( llm-stats; TechCrunch, Jan 28, 2026).

• Why it matters: both are leverage for on-prem / data-residency buyers and in price negotiations — pending independent benchmarks. Flag: vendor “most-capable/beats-Llama” claims; Trinity dated-older.

Meta’s Alexandr Wang says open-source is no longer sustainable for the most powerful models. He cited internal safety concerns around Meta’s Muse Spark model and said other labs face similar constraints ( Times of India, Jun 2026).

• Why it matters: a notable reversal from one of open-source’s loudest backers for buyers weighing open vs closed — though it’s also a vendor reframing its own strategy, so weigh it as such. Flag: vendor claim.

Other model updates worth tracking. BAAI claims the first general “world foundation model” understanding physical-world dynamics ( CGTN, Jun 14, 2026); Google’s Gemini 3.5 Pro is reportedly reaching GA this month alongside “Gemini Omnia” ( llm-stats); OpenAI’s GPT-5.5 remains the current frontier (Pro/Instant) ( felloai); Microsoft debuted its MAI models at Build ( llm-stats).

• Why it matters: BAAI is a geopolitical signal (China targeting physical AI) more than a benchmarked capability; the rest is model-selection context. Flags: BAAI vendor/state-media claim; Gemini GA timing reported-unconfirmed.