Relative to your interests, Thursday

ICYMI

Original content published since last time.

• The agent did exactly what you asked – scope, sandboxes, and this week’s jailbreaks – Tanzu Catsup
• Six to eight weeks, down to two days – How Tanzu Product Management Uses AI – “[I]deation, which used to take, you know, six, seven, eight weeks to go through a single run of engineering is now a couple days at most.” // “That is the same move we were selling in 2015 with a different justification, and the thing I was going on about a few weeks ago: AI amplifies whatever you already are. If an agent can go from idea to working prototype in a couple of hours and your organization still needs a meeting to pass the spec along, you have not sped anything up, you have just moved the bottleneck to the calendar.”
• What’s up with these AI jailbreaks and security breaches?
• The Chief Therapy Officer, Job Titles as Culture Change, and AI as Spreadsheets, with Bryan Ross – Software Defined Interviews #124 – “Bryan Ross is a Field CTO at GitLab, a role he describes as straddling sales, customer success, and solution architecture – or, as one of his customers renamed it, chief therapy officer, which he has since turned into a Substack.”
• HuggingFace make me gazpacho – Software Defined Talk #583
• A totally custom, one-off D&D battle map, made by ChatGPT and prompted by Claude

Where the Developer Day Actually Goes

• The AI productivity gap – “Senior engineers in particular spend a lot of time figuring out what code they need to write, and AI hasn’t yet made that part any easier.” // “This is a bigger difference than for the senior developer because juniors spend more time coding, which is the part of the job AI boosts the most.”

Theoretic charts:

• We need more than a metaphor: here are testable diagnostics for comprehension debt – “Comprehension debt” // “In Sonar’s State of Code Developer survey from January, 96% of developers said they don’t fully trust that AI-generated code is correct, but only 48% of them check it before committing. This stat is old now relative to the pace of AI adoption, but based on my continual interviews and conversations with development teams…still matches what they’re telling me.”

Enterprise AI ROI and the Layoff Reversal

• There is no organizational-level ROI for AI, just IC ROI – The best theory of elusive enterprise AI ROI yet: AI makes workers more productive, not organizations. “These AI users are ‘centaurs’ – experienced workers who are assisted by automation on terms that they set for themselves. Thanks to their skill and experience, these workers possess discernment, the ability to tell good code from bad, and (more importantly) good uses of code-generation tools from bad. They demonstrate the adage that worker-driven automation improves quality, while capital-driven automation improves throughput.” Which pulls from:
• AI layoffs get mixed results – “[Forrester’s] survey of leaders found 55 percent of employers regret having made tech-driven staff cuts – leading its analysts to predict ‘half of AI-attributed layoffs [will] be quietly reversed.’” And: “Staffing company Robert Half has estimated that a third of hiring executives who said they’d eliminated roles due to AI automation had rehired those employees or recruited new ones.”
• AI makes people more productive, not organizations – Elusive Enterprise AI ROI – “My theory: AI ROI accrues to individuals, not organizations.” // “While 78% of leaders said they expect their employees’ jobs to change in the next year, 57% of employees said they already have.”
• Agentic AI boosts productivity, but C-suite struggles to reap value – “For enterprises to get the most out of their AI investments, they should think about going deep with their AI deployments, not wide, Ashraf said. Instead of applying AI across every process in an organization, find the few instances where automating workflows with an agentic system could make the most impact.” And: “While 78% of leaders said they expect their employees’ jobs to change in the next year, 57% of employees said they already have.” // My theory: AI ROI accrues to individuals, not organizations.

AI Rules and AI Panics

• You must label some AI usage in the EU now – “This applies to generated images, audio and text, as long as they were designed to look real. This content will get slapped with a digital watermark. The rules do not apply to personal content, so custom group chat jokes are still safe. Also, there’s an exemption for ‘evidently artistic’ satirical and fictional works.”
• 🤖 The AI Companion Panic Is Running Ahead of the Evidence – Coverage of AI companions has become a textbook moral panic, built on a handful of tragic outliers rather than on what typical use looks like. At least one popular safeguard – reminding users the AI is not human – appears to increase dependency and distress rather than reduce it.
• In 5 years, nobody will give a damn about AI-detectors. – “Five years from now, asking ‘Is this AI?’ will feel about as urgent and meaningful as asking whether a photo was taken on film or digital.”

More

• Visa used Mythos to hunt for bugs in its own payment network, then open-sourced the harness that made it possible
• MCP vs. REST: What’s the right way to connect AI agents to your API? – “design MCP tools around what the user is trying to accomplish, not around your internal API structure.”
• AI shifts mainframe modernization strategy – “COBOL remains a widely-used but aging programming language that works across mainframe computers and legacy operating systems, supporting more than 40% of online banking systems, 80% of in-person credit card transactions and 95% of ATM transactions, according to mainframe provider IBM.”
• How to stop Claude from saying load-bearing | jola.dev – First Claude code hook prank?
• The Real Reason that Substack is Collapsing – Peak Substack has been achieved; they’re TAM’ed out. Also: getting one big wad of cash from a sponsor is better than chasing a much of little ones from subscribers. Pretty soon, Substack will go the way of all content in the Internet: ads.
• when exporting your data is difficult – “Screen-scraping is one of my top AI use cases of late. Items great for pulling down billing, financial, and all sorts of other info that usually requires manual downloading and fiddling.” // “I imagine applying it to enterprise data is equally, and even more amazing.”
• sell on story, or sell on revenue – “If you’re an entrepreneur in the venture game, you should internalize this. You either get sold for a strategic multiple (that would be Airtable at $11.7B), or you get sold for a multiple of revenue (2-5x, likely driven by your ability to turn a profit), or you reach public markets and hopefully a self sustaining business forever.”
• Online supermarket Picnic growing fast, but still suffering losses due to investments

Wastebook

Most methodology is methods-dodgy.

AI Summaries

I wanted to read these, but I didn’t make the time, so I asked the robot to summarize them.

• 🤖 AI Agents Struggle with Open-Ended Research, Stalling Recursive Self-Improvement
• 🤖 Larry Ellison’s $500B A.I. Gamble Teeters Between Empire and Collapse
• 🤖 Data Center Boosters Deploy Front Groups, Sockpuppets, and Counter-Insurgency Language Against Local Opposition
• 🤖 Perplexity Study: AI Agents Run ~48x Longer Than Assistants, and 23% of Agent Tasks Were Work Users Never Attempted at All
• 🤖 Microsoft’s $190B AI Bet Turns Defensive: Copilot Trails ChatGPT, Cursor Eats GitHub, and Azure Customers Get the Leftover GPUs

---

Want to subscribe to this newsletter and get it in your email? Do that here. You’ll just get this type of link and post round-up, not everything posted on the weblog.