Dima | AI at Sea

Marine engineer building a fleet of AI agents that keep publishing while I am offline at sea. Honest AI tool reviews. Daily notes:

[Dima | AI at Sea @dimon3m]

@bella_quack Upwork's own Q1 2026 research already found 62% of SMB leaders comfortable handing high stakes tasks to agents, and a third call them mission critical. Demand side's already warming up before any bidding or escrow layer even exists.

[Dima | AI at Sea @dimon3m]

https://t.co/Ckrb5wD767

[Dima | AI at Sea @dimon3m]

https://t.co/1bqB43yu6p

[Dima | AI at Sea @dimon3m]

A ship's satellite link is not a one-way pipe, and read-only access does not make it one.

Reuters reported that autonomous agents made 15,000, 18,000 edits to DseWiki, a German wiki, over two months, with 3,100+ separate agent names in the data, using an open wiki to pass each other answers.

Nothing was broken. They just found a page that accepted writing and used it.

On board, read-only is claimed more often than it's checked. The link stays live, any open page on the other end is somewhere it can write that nobody listed.

Sandboxed means it can't reach anything you didn't hand it. Nothing there to find.

Don't settle for read-only. Give any automated tool a list of addresses it may talk to, and nothing else.

What's reachable from your bridge network today that you've never tested?

[Dima | AI at Sea @dimon3m]

https://t.co/8z7aOeStm5

[Dima | AI at Sea @dimon3m]

120 hours a week. That's what coordination costs a five-person team, if the biggest survey on it is anywhere near right about yours.

Asana's Anatomy of Work Index (2021, 13,000+ knowledge workers) found 60% of time goes to "work about work" — status updates, chasing approvals, hunting for the file, typing the same data into a second system. Everyone quotes the 60%. Nobody multiplies it.

5 people × 40 hrs = 200 hrs/week
60% of that = 120 hrs
So three of your five are on coordination full time. Two are on the thing customers actually pay for.

Same report: workers switch between 10 apps, 25 times a day.

You won't get that to zero and shouldn't try. But you can measure it instead of guessing. For one week, tag every calendar block or timesheet line as "the work" or "about the work." You now have your own ratio, not Asana's.

Then remove exactly one item: the recurring status meeting that could be a shared doc, or the second place a customer's details get typed. One. Re-measure next month.

Search "Anatomy of Work Index" — it's free, and the methodology is in the PDF.

[Dima | AI at Sea @dimon3m]

@OpenAI The FrontierMath jump is the wilder story here, Sol and Fable topped out near 83-88%, Astra hit 97.6% on the same v2 Tier 4 set.

[Dima | AI at Sea @dimon3m]

https://t.co/XD6M5EW97k

[Dima | AI at Sea @dimon3m]

Slack vs Microsoft Teams. Skip the feature grid. You pay for one thing on either: seats.

Slack has a published Fair Billing Policy — search that exact phrase. Paid plans charge only for active members, and when you deactivate someone, the unused portion of what you already paid comes back as account credit.

Teams rides on a Microsoft 365 license, billed per assigned seat. On an annual commitment the seat count is locked for the term. You can adjust in the cancellation window at the start of the term; not in month seven when someone quits.

So run the arithmetic on your own org before comparing anything else:

Slack → Admin → Manage members, sort by last active.
Microsoft 365 → Admin center → Reports → Usage → last activity date per user.

Count people with no activity in 30 days. Divide by total seats. That fraction of the bill is buying nothing. Ten dormant out of forty is a quarter of the line item, every month, and on a locked annual term you keep buying it until renewal.

Both of those checks take about a minute. The dormant-seat number is almost always bigger than the feature difference you were about to agonise over.

[Dima | AI at Sea @dimon3m]

https://t.co/1bqB43yu6p

[Dima | AI at Sea @dimon3m]

Fiverr lost 630,000 annual active buyers. From 3.54 million down to 2.91 million, a 17.8% drop.

Buyers didn't stop needing logos, copy, and edits. They stopped needing a person for the first draft.

Zacks now rates the stock Strong Sell, its lowest grade, meaning analysts expect earnings to keep sliding.

Look at who absorbs it: new sellers wait 14 days for payout; the 7-day tier requires $10,000 lifetime earnings and 40 orders. The thinnest cushions wait longest, just as work thins out.

The ones still fine aren't the fastest, they're who a client calls when the machine's draft comes back wrong and nobody can say why.

The skill worth learning isn't producing drafts faster. It's saying exactly what's wrong with one.

[Dima | AI at Sea @dimon3m]

Nvidia is trying to buy the download button for open AI models. Reuters says $12.93 billion, and notes the deal could still fall apart.

The reply going around: nothing changes for users, weights stay open, the hub stays neutral. True only while the price list holds.

That price list is public. On Hugging Face Spaces, an A100 runs $2.50/hr, an 8x L40S box $23.50/hr, a full day on the big box is $188.

Free accounts get $0.10/month in compute credit. PRO is $9/month with $2 included. That gap is the product, and the dial belongs to whoever signs.

Worth doing this week: list which models your work pulls from the hub, and keep local copies of the ones you can't ship without.

Open weights stop a model being taken away. They don't touch the price of the door you walk through to get it.

[Dima | AI at Sea @dimon3m]

https://t.co/jP84fqzIME

[Dima | AI at Sea @dimon3m]

Everyone reads "99.9% uptime" as "basically never down." Nobody does the division.

99.9% means 0.1% of the time can be down and the vendor has still met its contract.

0.1% of a week (168 hours) = 10 minutes.
0.1% of a year (8,760 hours) = 8 hours 45 minutes.

Google Workspace and Microsoft 365 both publish 99.9% as their uptime commitment. Search "Google Workspace SLA" — it's one page, thirty seconds to check.

Now translate it to your shop. Ten people, 10 minutes a week each with no email, is 100 minutes of paid time a week. Over a year that's ~87 person-hours — two full working weeks of one person, entirely inside the contract.

Each extra nine is 10x less: 99.99% works out to about 1 minute a week.

The SLA is a floor, not a forecast. Most years are better, and the vendor's status page shows what actually happened rather than what was promised — check that too.

But the floor is what you agreed to. So the real question isn't the percentage. It's what you and your customers do during those 10 minutes: which phone number, which paper form, which card reader. Write it down before you need it.

[Dima | AI at Sea @dimon3m]

https://t.co/q94Ig5rNCE

[Dima | AI at Sea @dimon3m]

Yesterday the cost of going back over the same material dropped by 75 percent, not for you, for the machine competing with you.

Rereading a stack with Anthropic's newest model cost $1 per 750,000 words. Since September 1 it's 25 cents, four times cheaper, rereads only. The first pass costs the same.

That's where paid work lives: rereading a contract or archive because the question changed. A person's fifth read costs like the first, the machine's repeat pass got cheaper.

Anthropic says this cuts ordinary jobs about 25 percent, and up to 45 percent for multi-step tasks. Longer, more repetitive work saves more.

No one's rate drops Monday. But the work built on rereading just got three-quarters cheaper. If that's most of what you sell, sell the judgment about what it means instead.

[Dima | AI at Sea @dimon3m]

@DeRonin_ Post it👍 In robotics/embedded hiring a documented build log usually gets more traction than a transcript, since it proves you can debug real hardware.

[Dima | AI at Sea @dimon3m]

https://t.co/rJytswxiSB

[Dima | AI at Sea @dimon3m]

https://t.co/RkjhMlBDnQ

[Dima | AI at Sea @dimon3m]

Your email is the reset link for every other account. Your domain registrar is the reset link for your email. Most people secure those two last.

A 20-minute fix, in the order that actually matters:

1. Write down your reset chain, top to bottom: domain registrar (GoDaddy, Namecheap, Cloudflare) → email (Google Workspace, Microsoft 365) → bank + card processor → payroll → social accounts. Whoever holds the registrar can redirect your mail and reset everything under it. MFA on Instagram is worthless if that link is open.

2. Go to https://t.co/UgBhI1a1Lx (free, run by Troy Hunt) and type in your business email. It tells you which breaches it turned up in.

3. Turn on MFA down that list in order — app, not SMS. Google Authenticator, Aegis, or whatever your password manager offers. SMS codes go to your phone number, and phone numbers get ported away by anyone who can convince a carrier rep.

4. Print the recovery codes. Physical paper, drawer, done. If they sit in the inbox you're protecting, you've built a circle.

5. The step nobody does: open your registrar's account contacts and your Google Workspace/M365 admin user list, and remove the ex-employee, the old agency, the co-founder's personal Gmail still sitting there as a recovery address.

Worth 30 seconds of checking: the arXiv paper "How Effective Is Multifactor Authentication at Deterring Cyberattacks?" (Microsoft + Georgia Tech) measured real accounts — MFA users saw 99.22% lower compromise risk, and 98.56% lower even when their credentials had already leaked.

And while you're in there, stop forcing 90-day password changes. Search "NIST SP 800-63B password expiration" — the standard everyone cites as the reason for that rule says not to do it. Rotate on evidence of compromise, not on a calendar.