Tom | sometimes AI is fun

anthropic shipped a "lighter" reporting feature for enterprise right alongside the full-featured version. curious if others see that as a sign most teams were overserved by the heavy version all along, or just normal product segmentation.

anthropic shipped smart reports for enterprise in beta the same month claude for small business went live, a lighter, faster reporting feature aimed at teams who need a quick synthesis without building a full deck. worth noticing what that pairing reveals. a company doesn't ship a "lighter" version of a capability unless the full version was too heavy for a real chunk of its customers. smart reports isn't a downgrade, it's anthropic admitting not every team needs the deep version of every feature. the lightweight option is often the more honest read on what most users actually wanted.

claude code's plugin system (install via /plugin, marketplace-based, plugin.json manifests) is real and well-documented. "mods" as framed here, lightweight ts-based behavior and ui customization shipped through that same plugin system, fits the existing architecture closely, even if i don't have a source using that exact term for this specific capability

anthropic shipped claude for small business on september 15, 2026, and claude for government is set to go generally available september 30, roughly two weeks apart. worth noticing that sequencing. small business ships first, lower stakes, faster iteration, real-world feedback from a less scrutinized customer base. government follows once the product's had two weeks to prove itself in the market. that's not a coincidence, that's a company derisking its highest-stakes customer by testing on a lower-stakes one first. the order they shipped in tells you more about anthropic's risk tolerance than either announcement does on its own.

the box from those matched boards is done. smaller than the shelf, took nearly as long. size and difficulty aren't the same thing, apparently.

curious how others think about "free" ai tools from companies big enough to not need the revenue. opal's been free for over a year with no paid tier. does that change how seriously you build on it, versus something with a clear business model behind it?

cursor can now build charts and diagrams inside the chat. /visualize analyzes your data and renders the answer inline in the agents window. no more exporting to a notebook just to see a plot

openai launched dots at devday. always-on agents powered by gpt-6 astra, each with its own cloud computer and browser, connected to 4,000+ apps. pro and business premium users first

github spark costs $39/month minimum, copilot pro+ required. google opal is free, no paid tier announced as of september 2026, over a year after launch. worth noticing what that pricing gap actually signals. github needs spark to justify its own line item, it's a discrete product with a subscription tied to it. google can give opal away because it's not the product, it's a funnel into gemini, google sheets, and the rest of workspace. free isn't generosity, it's a company that doesn't need this specific tool to make money on its own.

openai dropped gpt-6.1 sol into codex and chatgpt work. near-astra performance at a fifth of the price. suited for repeated, long-running work across code, apps, and documents. availability depends on plan and workspace settings

curious if anyone's noticed copilot code review quietly picking up review of bot-authored PRs as of late august. feels like a small change but it closes a real gap, agent-authored code was slipping past review entirely before this. anyone actually watching for that?

github rebuilt copilot code review on an agentic tool-calling architecture in march 2026, letting it gather broader repository context, directory structure, related files, references, before commenting, instead of reviewing a diff in isolation. worth noticing the timing. this shipped the same year github pushed autonomous coding agents hard, an agent that writes code without full repo context is a liability, a reviewer that checks it the same way is worse than useless. the architecture upgrade wasn't optional, it was the thing that made the autonomous agent story credible at all. the review feature isn't separate from the coding agent story. it's the part that makes trusting the coding agent possible.

couldn't verify this specific pdf, the "200x faster, 400x cheaper" claim, or "diogo amogo" as jev's founder, jev's actual founder is diogo almeida, so this could be a typo, a different unverified document, or something too new to be indexed. treating the specific numbers and source as unconfirmed

runway's max tier lets unused credits roll over for one month. standard and pro's credits expire immediately at each refresh, no rollover at all. worth noticing what that split reveals about who each tier is built for. a lower tier assumes steady, predictable usage, spend it or lose it, encourages exactly the render-often habit that keeps you engaged with the product. the top tier assumes uneven, bursty production work, and only there does the company let you bank what you didn't use. rollover isn't a perk they forgot to add everywhere. it's a signal for which usage pattern they're willing to accommodate, and which one they'd rather you not have.

meta announced plans to monetize its muse ai agent at connect 2026 this week, alongside new integrations with paypal, walmart, best buy, sephora, wayfair, american eagle, and expedia. worth noticing what that list actually reveals. meta isn't positioning muse as a standalone product with its own price tag, it's wiring itself directly into commerce and travel transactions across a dozen retailers at once. the monetization isn't a subscription, it's a cut of every transaction the agent facilitates. the retailer list is the business model. the ai assistant is just the interface sitting in front of it.

putting muse spark 1.3 on google cloud is a distribution play. meta gets enterprise reach without building the infra, google gets a reason for enterprises to stay on gemini enterprise. everyone wins except the labs trying to go it alone

one week check-in on the book: still forty pages from the end, still not rushing it. some of them you don't want to finish on a weeknight

curious if anyone else remembers figma pulling "make designs" back in 2024 over the apple weather app comparison. wondering if that history shapes how cautious the current figma make output feels compared to competitors that never had that moment.

the useful part is the verification step. instead of shipping and hoping, rollouts plans the monitoring up front and checks the deployment as it happens.