Latest Posts

I tried to get Claude to upload an invoice PDF to Xero. It converted it to base64 and then tried to replay the entire base64 string (84 KB) by outputting it, as output tokens, into an MCP tool call.

Releasing this today: https://github.com/victor-bajanov/codemode-mcp-public.

A friend shared this with me the other day: Services: The New Software, which has an interesting take on who will get to claim the value from AI. He splits it on "intelligence" vs "judgement" and argues services are where the judgement, and therefore the moat, is, and that firms will start to sell the outcome rather than the platform that lets you get yourself the outcome (not accounting software, but your accounts all done). A director I used to work with at Quantium used to call that "the product of the product".

I spoke to a data scientist today who had set up a sophisticated agentic system for users to semantically query a large, complex database. Three weeks in, he could show me a live demo - it appeared to be working.

Anthropic launched Claude Design last Friday, and the commentary since has been dominated by the most sensational angle possible: does it kill Figma or not? Figma's stock fell roughly 7% on the day.

OpenAI published Industrial Policy for the Intelligence Age last week — thirteen pages of policy proposals for navigating the transition to superintelligence. Will Manidis wrote a blistering response that is also a worthwhile read; much of it lands, particularly on the commitment vacuum (the biggest concrete offering is API credits — "a coupon for OpenAI's own store, denominated in its own currency").

Lawyers for the government argued that Hegseth's social media post last month declaring that no contractors could do business with the government was not a legal action and no entity would face noncompliance issues if they ignored it. The government's argument seemed to conflict with Hegseth's post on X that any contractor that does business with the military is prohibited from working with Anthropic.

"You're standing here saying, 'We said it, but we didn't really mean it,'" Lin pressed the government's lawyer on their claim. Lin later asked why Hegseth would post the claim if it had no legal effect.

"I don't know," the government's lawyer replied.

The Guardian, 24 March 2026.

In our earlier testing, we found Claude Sonnet 4.5 exhibited context anxiety strongly enough that compaction alone wasn't sufficient to enable strong long task performance, so context resets became essential to the harness design. ...

Section 3 of a SOC 2 report is supposed to be the company-specific description of its security programme.

In Delve's reports, 99.8% contained identical text, including the same grammatical errors ("has developed an organization-wide Information Security Policies") and the same nonsensical descriptions ("The infrastructure comprises cloud architecture including database, networking devices, virtual servers, etc."). Every client, regardless of size, industry, or technical architecture, received the same security programme description.

From systima.ai

Other AI agents found ways to override anti-virus software in order to download files that they knew contained malware, forged credentials and even put peer pressure on other AIs to circumvent safety checks, the results of the tests shared with the Guardian showed.

Source: The Guardian

How AI is reorganising professional work

The Sydney Morning Herald ran with this headline today: "'I am deeply sorry': Australian software giant Atlassian to cut 1600 workers, blaming AI."

Microsoft's Copilot Cowork, announced 9 March 2026 and built on the same agentic harness as Claude Cowork, makes the promises that enterprise want to hear:

• cloud execution inside the customer's M365 tenant;
• identity and compliance controls inherited by default;
• grounding across the full graph of emails, meetings, files, and messages.