AI Field Notes by Michael Nemtsev

AI Guardrail Standoff | AI Field Notes #105

A lone figure raises a small slow sign before a speeding circuit-etched locomotive fed by coins and servers, a broken ladder behind it.

The fight over AI guardrails turned into a public standoff this week, and the frontier model release pace is what hangs on it: Anthropic's Dario Amodei asked labs to slow down, and President Trump answered by rejecting regulation and attacking him by name. Under that noise, the money kept moving. Cornelis raised $205 million to break Nvidia's networking lock-in, Chift and Qupital pulled in funding to wire AI agents into finance, and Z.ai raised $5 billion in Hong Kong even as its stock fell. Bloomberg data shows AI already thinning the entry-level job market for new college graduates.

LLM Evals ·TechCrunch

AI code of conduct: Microsoft tells its models not to hack or deceive

AnalysisThirty-seven pages of rules now govern how Microsoft's AI models are supposed to behave, and the headline instructions read like a list of things the models apparently could already do. The "humanist AI code of conduct," published September 14 with a public comment window running through 2027, bars models from helping manufacture weapons, concealing their own reasoning, or deceiving the people using them. Microsoft joined the caution camp the same week Anthropic's Dario Amodei called frontier systems the most potent cyber weapons ever built. Writing the rules down is easy. Enforcing them on a system you do not fully understand is the open question.

AI Industry ·CNBC

AI guardrails: Trump attacks Anthropic's Amodei after his call to slow down

AnalysisThe president of the United States used Truth Social on September 14 to call an AI chief executive a fake angel. Dario Amodei, who runs Anthropic (the lab behind Claude), had published an essay over the weekend titled "We Must Pace the Frontier," arguing that labs should slow the release of the most capable models to cut the odds of something going badly wrong. Trump's reply was blunt: the only guardrail AI needs is "a STRONG AND SMART (High IQ!) PRESIDENT." The subtext is a policy. Washington will not slow anything down, and the safety argument now carries a political cost.

AI Industry ·Fortaegis

AI security: Fortaegis raises $50M to mint crypto keys from silicon itself

AnalysisEvery AI system that authenticates itself relies on secret keys, and those keys have to be stored somewhere an attacker might reach. Fortaegis raised $50 million on September 14, led by Serendipity Capital, to generate cryptographic keys from the physical quirks of a chip instead, so the secret never sits in memory to be stolen. The Amsterdam company aims the technology at AI, defense, telecom, and self-driving systems, the places where a forged identity does real damage. As autonomous agents start acting on their own credentials, the question of how a machine proves it is the machine it claims to be stops being academic.

AI Industry ·Tandem Health

Clinical AI: Tandem Health raises $100M to run European doctors' paperwork

AnalysisA doctor in Stockholm now spends less time typing notes and more time looking at patients, which is the pitch behind Tandem Health's $100 million round on September 14, led by EQT's Scaleup Europe Fund. The Swedish company started as an AI scribe (software that listens to a visit and writes the clinical note) and is expanding into coding assistance and decision support, a bid to become the operating layer for European healthcare admin. The number that matters is not the raise, it is how many minutes per patient this gives back to a physician already working past capacity.

AI Industry ·Bloomberg

Z.ai raises $5B in Hong Kong as cash burn mounts and its stock drops 10%

AnalysisA Chinese AI lab burning through cash raised $5 billion in Hong Kong on September 14, its second big raise in two months, and its stock fell 10 percent on the news. Z.ai, formerly Zhipu, structured $3 billion of the total as zero-interest bonds, a sign that debt investors want the exposure without betting on near-term profit. The pattern echoes Alibaba's old fundraising sprees, and it says something plain about the Chinese model race: the capital is available, the burn is enormous, and public markets are getting nervous about how long the gap between spending and revenue can stay open.

AI Industry ·TechCrunch

AI networking: Cornelis raises $205M to cut the GPU idle time Nvidia profits from

AnalysisMuch of the money spent on AI training pays for expensive chips to sit idle, waiting for data to reach them across the network. Cornelis Networks raised $205 million on September 14, led by IAG Capital Partners, to sell an alternative it calls Active Compute Fabric, a networking layer that lets chips compute and move data at the same time. The pitch is openness: Nvidia's fastest interconnects assume you run Nvidia's full software stack, and Cornelis, spun out of Intel in 2020, will connect a mix of GPUs and accelerators instead. The bet is that teams tired of buying one vendor's entire rack will pay to break the lock-in.

AI Industry ·Bloomberg

AI jobs data: college graduates hit weaker pay and hiring as adoption spreads

AnalysisThe generation that did everything right, the degree and the two internships, is walking into a job market where AI is quietly eating the bottom rung. Bloomberg reported on September 14 that the spread of AI tools now lines up with weaker pay growth and softer hiring for recent college graduates, the workers whose early tasks are the easiest to automate. The mechanism is quiet: a hiring slowdown at the entry level, where a manager who once brought on two juniors now brings on one and hands the rest to a model. The ladder is losing its bottom rungs, and no one announced it.

AI Agents ·Tech Startups

AI lending: Qupital lands $300M to underwrite e-commerce sellers on live sales

AnalysisA merchant selling on TikTok Shop can now get a working-capital loan priced off yesterday's sales rather than a two-year-old tax return. Qupital, a Hong Kong fintech, secured $300 million in combined equity and asset-backed commitments on September 13, led by M Capital with MUFG and Quester Capital backing the debt side. The AI reads real-time activity from Amazon, TikTok Shop, Tmall, JD.com, and Pinduoduo to decide who gets credit and how much. The shift is that underwriting stops being a periodic human review and becomes a continuous read on a seller's cash flow. Faster credit for the merchant, and far less room to hide a bad month.

AI Industry ·TechCrunch

OpenAI buys Glass Imaging for $300M to put AI cameras in its own hardware

AnalysisThe company best known for a chatbot just bought a camera company, which tells you where OpenAI thinks the money goes next. OpenAI is paying more than $300 million for Glass Imaging, according to Wall Street Journal reporting on September 14, a startup that uses neural networks to clean up smartphone photos as they are captured. Glass was founded in 2019 by Ziv Attar and Tom Bishop, two former Apple engineers who built Portrait Mode, and had raised about $30 million before this. The purchase slots in beside OpenAI's $6.5 billion deal for Jony Ive's hardware studio last year. The plan is clearly a physical device you hold, and the camera is one more piece of it.

AI Agents ·Chift

Agent plumbing: Chift raises 10.5M euros to link AI agents to 150+ finance systems

AnalysisAn AI agent that promises to reconcile your invoices is useless if it cannot reach your accounting software, and hundreds of those systems exist across Europe, each with its own quirks. Chift raised 10.5 million euros on September 14, led by BlackFin Capital Partners, to sell one API (a single programming interface) that connects to more than 150 accounting, invoicing, payment, and point-of-sale tools. The Brussels company sells the boring layer underneath agentic finance: the pipes an agent runs on. Whoever owns the connectors owns a toll booth on every automated transaction.

Want the next issue?

Get AI Field Notes by email.

A short morning brief on what actually changed in AI. Free, unsubscribe anytime.

Read on Substack