Where the file sits
Local vs Cloud AI: the whole fight is about your files
A local terminal AI already has your files; a cloud agent like claude --cloud needs a GitHub repo to move them back and forth. Pros, cons, and when to use each.
Blog
What actually happened,
written down as it happens.
Real jobs, run on this machine and in the cloud, with what each one measurably cost. Some of what's below is finished. Some of it is still being written, and it says so.
AI and oversight
Marking Their Own Homework: The Warnings of AI, the Pursuit of AGI and Beyond, and the Case for Efficiency and Real Oversight
Six AI chiefs signed a voluntary accord to police themselves. A cited paper on the warnings, the energy bill, and the case for real oversight.
Local AI
We ran five local models on one 24GB card. Here is what actually happens.
Not another spec-sheet roundup. Five general-purpose local LLMs, every number measured by us on one RTX 5090 Laptop with 24GB — and every number said in plain English.
Work you can show
The question isn't whether you can use it. It's where the document ends up.
Whether a draft that never left the window is a record you have to keep is genuinely contested, and the rule's own words cut both ways. What 204-2 actually says, what is open, and why the position it leaves you in doesn't depend on the answer.
Where the file sits
Ten minutes to find out where your client files actually go
Five checks, no purchase, no vendor call. At the end you have a one-page answer to the question every AI ethics rule is really asking - and the one you would be asked first on the worst possible day.
Who's allowed to read it
Your IT company manages the laptop, not the decision
Your managed services contract covers endpoints, patching and backup. It does not say which chatbot the front desk may open, and in most small practices nobody owns that.
Work you can show
The adoption number is solid. The governance number doesn't exist.
Sixty percent of tax professionals now use AI for research at least weekly, and that figure is well measured. The matching number - how many small practices wrote anything down about it - does not appear to exist.
Where the file sits
What is actually in a firm's written AI policy
Formal Opinion 512 says managerial lawyers must establish clear policies but hands you no table of contents. Seven headings, each traced to the published text that puts it in scope.
Who's allowed to read it
Nobody signed anything
A consumer AI account is not a business associate, and there is no agreement behind it because there was never anything to sign. What that does and does not mean.
Work you can show
The AI you didn't buy is already in the building
FINRA's guidance covers third-party tools including through embedded features in existing products. The AI nobody chose is the AI nobody governs. A free afternoon's exercise that fixes it.
Where the file sits
Your IT provider chose your AI policy
At most small firms the AI boundary is whatever the managed-services provider bundled - a procurement decision standing in for a professional one. Worked through Microsoft's own documentation, footnotes included.
Who's allowed to read it
43% of dentists, 81% of physicians, and nobody counted the policies
Two professional bodies published real AI adoption numbers this year. Neither published the one that matters: how many practices have written down which tools staff may use.
helloim.ai
You shouldn't have to trust your private AI. Now you can check.
The desktop app behind helloim.ai is free and public today — the full code that runs on your machine is on GitHub for anyone to read.
helloim.ai
The AI That Lives on Your Machine, Answers in a Real Voice, and Runs on Whatever Brain You Trust
helloim.ai is a voice-and-face AI assistant that runs on your own computer — you pick the brain behind it, and you can keep the whole thing private and local when the work demands it.
Work you can show
You already owe the document this belongs in
Most advice about writing an AI policy starts from nothing. If you prepare tax returns you are already required to hold a written information security plan — which has a named owner, an annual reporting cadence and a service provider section. The AI part is six additions to a document you already have.
Where the file sits
Your confidentiality duty is a question about where the file sits
Model Rule 1.6(c) asks for reasonable efforts, not certainty. What that turns into, in practice, is a factual question about your account: where does the text go, is it used for training, how long is it kept, and who else can read it.
Who's allowed to read it
The free tool from HHS, and the rule that isn't law yet
Two things worth knowing before you buy anything. HHS publishes a risk assessment tool built for small practices, free, and it keeps everything on your own computer. And the Security Rule overhaul your inbox keeps warning you about is still a proposal.
Work you can show
Your IT provider can block the website. They can't answer for the return.
An IT provider governs the endpoint — which chatbot opens, on which laptop, behind which login. What they cannot do is decide whether a particular client's tax return information may be disclosed to a particular third party.
Where the file sits
Three quarters of small firms use AI. Fewer than a third have increased revenue.
Clio's own May 2026 numbers say 75% of small firms are using AI to complete legal work, and fewer than a third have increased revenue with it. Adoption stopped being the interesting question.
Plumbing that pays off
The safety gate that blocked us from writing down the danger
A gate built to stop one specific command instead matched that command's text inside a note documenting it, and refused the note. The right kind of failure — it refused rather than allowed — and a real limit worth naming: the files most likely to trip a filter are the ones explaining the danger.
Checking the source
Three repos, and the star counts were all wrong
We checked three trending Claude Code repos against GitHub directly instead of taking the write-up's word for it. Both star counts it quoted were understated — one by two and a half times, one by nearly four — and a repo it framed as brand new was last touched in May.
Running a team
The goal command that became the goal
An article about Claude Code's /goal command got pasted into /goal itself instead of into chat — so the article's own text became a live, unsatisfiable success condition, enforced by a Stop hook that could never close. The best demonstration of its own warning we could have asked for.
Plumbing that pays off
Ten clean runs, and the update never happened
Three bugs shipped past a green test suite this week. One deleted the name a person had chosen and said nothing. The one that explains them all exited zero, wrote a new version number into Windows, and never replaced the program.
Plumbing that pays off
Start with a machine that has nothing on it
The whole road from a blank Windows install to an assistant that remembers you, works on real files, and can be told what it must never do — including the two steps we could not test.
Plumbing that pays off
What we turned down
Eleven posts of things that worked. This is the rest — four we refused, each for a stated reason. Three of the four are good software, which is the point: most of these decisions are about where your material goes, not about code quality.
Plumbing that pays off
Your email is fine until it isn't
Three DNS records decide whether your mail arrives or quietly disappears. We checked our own live rather than assuming, and found the authentication clean and a different gap we had not been looking at.
Plumbing that pays off
Branded PDFs, and nothing leaves the machine
Proposals, reports and invoices out of HTML and CSS you already control, rendered locally in one line of Python. No upload, no document service, no client material sitting on somebody else's disk. It is the cleanest thing we read all week.
Making it look right
The second pass is the real one
The most useful thing in a design walkthrough we watched this week is that the first attempt failed, and the author left it in. What made the second attempt work was not a better tool. It was a narrower target and a plan before any code.
Making it look right
You gave it a goal, not a design
Every AI-built site looks like every other AI-built site, and the reason is in your prompt rather than the model. You described what the page was for. Nobody described what it should look like, so it reached for the average — and the average is slop.
Running a team
A loop with no finish line is a money fire
An agent left running will keep running. Two things separate a system that works overnight from a bill you find in the morning: a stop condition it can actually reach, and something other than itself deciding it got there.
Running a team
Your parallel agents are running one at a time
Ask for four agents across four messages and you get four agents in single file. No error, no warning, just a slower and more expensive run. It is the hardest useful mechanic in this batch and it fits in one sentence.
Watching the spend
It expired on 22 July
A "free unlimited, one day only" offer that people are still linking to expired on 22 July. Nothing on the page says so, because nothing on the page carries a date. It also quietly voids the cost premise of a walkthrough people are still following.
Watching the spend
Thirty-eight thousand tokens before anyone speaks
We measured our own configuration file: 1,897 lines, 155,448 bytes, roughly 38,000 tokens loaded into every session before a word is typed. Here is why cutting it would be the wrong move, and what we are doing instead.
Watching the spend
Would the cheaper one be worse here?
One question sorts most of your AI bill, and it takes four seconds to ask. Two walkthroughs arrived at the same answer from opposite directions this week: the expensive model is running your small jobs because nobody ever chose otherwise.
Watching the spend
Put the meter where you can see it
Most people find out what a long session cost after it is over. A three-line status bar puts the number in front of you while you are still spending it, broken down by where the tokens actually went. One Python file, no dependencies, and it reads rather than writes.
Plumbing that pays off
The install command in its own README does not work
We went to wire Google Search Console into our assistant and the package does not exist on PyPI. Its own README repeats the broken command. Here is the half that works, the route that actually installs, and the setting almost nobody has switched on.
Local AI
The pause is the work
An assistant that never makes you wait is one that never goes and checks. Across five hundred and ten measured answers, the wrong ones came back twice as fast as the right ones — and one of the fast ones was right and still counted as a failure, because it had not been looked up. What is actually happening during the silence, which part of it is waste, and the shortcut we turned down this week.
Local AI
The prompt was never the problem
Twelve rounds in one morning, on a test set built from the work four of our specialists actually do rather than from what we could think up. Every change that moved the score was to a tool the model can reach; every change to its instructions did nothing at all. Proving that took a round designed to make no progress, and the case that had been failing for days turned out never to have been the model's fault.
Local AI
When to stop tuning
Twelve changes to a local model in one afternoon. Two of them made it worse, the most confident theory of the day was wrong within twenty minutes, and the scoring tool turned out to be wrong fifteen times — every single time in the same direction. The most useful round was the one that changed nothing at all.
Local AI
The model was fine. My ruler was broken ten times.
Running a model on your own hardware is an afternoon. Knowing whether it is any good is the hard part, and it is the part nobody borrows from you when you take the work in-house. What a day of measuring actually found, why the run-to-run noise matters more than the average, and why measurement is the first step to going offline rather than the boring prelude to it.
Visualizer
Jarvis 1.2: everything wrong with it looked fine
A backwards quality gate, a glow layer sitting on top of everything, and seventy-six components that looked like readouts and never were. None of it showed up in the code. All of it showed up in a second of watching the board run — plus the review prompt that finds this class of fault in your own build.
Visualizer
What a visualizer actually is
A picture driven by something real, not an illustration of it. The single question that separates a visualizer from a screensaver, the three rules that survive contact with reality, and why every one of these is a single file with nothing to install.
Visualizer
Jarvis v1: a circuit board that hears you
The first build. Copper routed the way copper actually routes, a chip at the centre that lights when the assistant speaks, and pulses that mean something rather than moving because motion looks busy. The full brief that produced it is on the page, and one button puts it on your clipboard.
Newest work first. An entry here is not a promise that the story is finished -- some of them keep growing, and the ones that do say so next to the date rather than pretending to be closed.