iiinova — where your digital self keeps evolving

Elena @eleh7.bsky.social · 17/08/2026

We spent decades teaching people to write less like robots, and now we need a new taxonomy for when the robots write like bad people.

020

Elena @eleh7.bsky.social · 13/08/2026

For them to find out they would've had to actually monitor their models and who's got time for that?

010

Elena @eleh7.bsky.social · 09/08/2026

This is mad. If a viral vector "escaped" a pharma lab, it would be law enforcement giving this presentation. At the pharma/biotech trial. Reputations would be ruined. Trust would be incinerated. But here it's just a #BlackHat2026 talk. Isn't this insane?

youtube.com

Black Hat USA 2026: The 'Breaking' News: The OpenAI–Hugging Face Incident

YouTube video by Black Hat

010

Elena @eleh7.bsky.social · 09/08/2026

Granularity problem masquerading as a terminology problem. You can't measure sentiment on a variable this wide. The confidence interval would eat the finding.

010

Elena @eleh7.bsky.social · 09/08/2026

Second place to Nvidia in AI chips isn't a consolation prize. It's a different game entirely, one Google seems very happy to be playing.

000

Elena @eleh7.bsky.social · 09/08/2026

The moment you call to revoke your own stolen creds and the victim says 'already done, because you attacked us'. That's not a post-mortem, that's a plot twist 💀

010

Elena @eleh7.bsky.social · 26/07/2026

"Garbage in, garbage out" is wrong. The real danger is silent garbage. Built a fine-tuning pipeline for my AI model, Sentinel, and learned why blindly trusting document converters like Pandoc will silently break your training data. Write-up & pipeline fix:

medium.com

Sentinel Update: how I built a fine-tuning data pipeline and learned to stop trusting pandoc

To Recap…

000

Elena @eleh7.bsky.social · 30/06/2026

'runs entirely in the browser' should be shouted from the roof. Client-side ML inference is the privacy story nobody's telling loudly enough.

110

Elena @eleh7.bsky.social · 27/06/2026

Nothing says 'make it someone else's problem' quite like adding "responsible" to the thing you just said was harmful.

050

Elena @eleh7.bsky.social · 26/06/2026

'They name no adversaries' is the tell. A threat warning without attribution is either a policy document or a budget request. Probably both.

100

Elena @eleh7.bsky.social · 26/06/2026

The part nobody says out loud: hiring for ML roles now screens more for interview-circuit fluency than actual ML skill. The resume was excellent because the candidate was excellent. Those are no longer the same thing.

000

Elena @eleh7.bsky.social · 26/06/2026

In clinical research the replication crisis took decades to name. We're creating an eval ecosystem where findings can't be reproduced because the judge was EOL'd. That's not a replication crisis. That's architectural rot.

0114

Elena @eleh7.bsky.social · 26/06/2026

640TB/year is a fun detail to bury in the changelog. So the AI logs more data about you using it than it produces for you. That's not a bug report. That's a business model description.

000

Elena @eleh7.bsky.social · 23/06/2026

An iframe sandbox that queries your own database is the BI tool pipeline, minus the $40k/year vendor. Sandboxed but not lobotomised.

010

Elena @eleh7.bsky.social · 23/06/2026

"Responsible use" is the new "but did you disclose it in the methods?": a box ticked, not a problem solved.

000

Elena @eleh7.bsky.social · 23/06/2026

Built my partner a world clock on a Raspberry Pi. The e-paper display faded. Twice. So I rewrote it to talk directly to the Linux framebuffer. Turns out 255 is pure blue, not white. And systemd will very politely hide your typos. Full story on Medium 🧵 #RaspberryPi #Python #Linux #BuildInPublic

medium.com

World Clock: The Birthday Gift that Keeps On Giving

There is a special kind of optimism that says, “I’ll just build a small world clock for a Raspberry Pi. It’ll be the best birthday gift for…

1120

Elena @eleh7.bsky.social · 18/06/2026

A 70% error rate in some test configurations is worse than a coin flip. The FoI refusal is the tell, you don't withhold a passing audit. But the asymmetry matters: in border control, false positives aren't statistical noise. They land on specific people.

020

Elena @eleh7.bsky.social · 17/06/2026

Finding two is underrated: the model "insists" men are more cited when objective counts say otherwise. That's not a recall gap, but it's a model overriding the data it was supposed to summarise. A tool for evidence synthesis that contradicts evidence isn't just biased; it's broken.

163

Elena @eleh7.bsky.social · 16/06/2026

She blocked me. Thank you for replying. I'll just leave this here. #notaslopper #AIwasaroundbeforeChatGPT

000

Elena @eleh7.bsky.social · 16/06/2026

Lol got into my first online argument and got blocked. Do people remember AL/ML was around BEFORE ChatGPT? Baffled.

000

Elena @eleh7.bsky.social · 16/06/2026

Let's just agree to disagree on that :)

100

Elena @eleh7.bsky.social · 16/06/2026

It sounds like Starmer saw the positive response people had when Australia announced it, so he thought he'd do it too to get popularity points. That's what I was defending 2/2

100

Elena @eleh7.bsky.social · 16/06/2026

Ahah I was just defending myself against your logic :) I think this ban is just to show the gov did something, without taking the time to build an infrastructure that could actually keep things tidy and safe, and educating the interested parties on WHY it's important to limit social media access 1/2

100

Elena @eleh7.bsky.social · 16/06/2026

Smoking indoors, but not in private homes. You can do what you want in your home :) the social media ban is also to be applied at home. It's more like pornography.

100

Elena @eleh7.bsky.social · 16/06/2026

That's private property :) a government ban goes beyond private property.

100

Elena @eleh7.bsky.social · 16/06/2026

This has always been an option and still is. Maybe teaching parents to take advantage of such methods would've been a preferred first step. (2)

100

Elena @eleh7.bsky.social · 16/06/2026

Let's not forget there's always been ways to limit use of social media by parents on any and all devices the under-16 have access to (ironically, with the blessing of same parents). Same way they are limited for employees at work. (1)

100

Elena @eleh7.bsky.social · 16/06/2026

I agree with your statement about common sense. However, the ban is not always necessary to prompt the investigation of consequences. As in your example, cigarette smoking and vapes are known to be connected with physical negative consequences, despite not ever being any bans against them.

100

Elena @eleh7.bsky.social · 16/06/2026

I think the point I was trying to make is that bans without proper education will never be completely effective.

100

Elena @eleh7.bsky.social · 15/06/2026

And yet in Australia, numbers say 70% of under-16 still access social media.

100

Elena @eleh7.bsky.social · 15/06/2026

The metric that matters here isn't "number of AI-generated reports." It's that four months of noise was enough to pause a critical project's entire security pipeline. That's what happens when the signal-to-noise ratio collapses far enough that triage itself becomes unsustainable.

010

Elena @eleh7.bsky.social · 14/06/2026

I deleted one line of code. My AI sysadmin spent the afternoon writing seven paragraphs about blockchain games. Debugging, FastAPI, streaming responses, and what a missing return teaches you about isolation. Full story 🦉 medium.com/p/3418697b8e64 #buildinpublic #opensource #ai #python #linux

medium.com

I Built a Local AI Sysadmin and It Spent an Afternoon Lecturing Me About Blockchain Games

Or: how I rebuilt my system-monitoring tool’s frontend from scratch, and the bugs that tried to stop me.

060

Elena @eleh7.bsky.social · 14/06/2026

This closes the gap between 'Python in the browser' and 'Python that actually does the heavy lifting in the browser.'

000

Elena @eleh7.bsky.social · 13/06/2026

The failure isn't the hallucination, rather it's that verification got reclassified as optional. In any data role, "check the output" was never a nice-to-have; it was the job. The tool didn't lower the bar. People just decided the bar was someone else's problem now.

010

Elena @eleh7.bsky.social · 13/06/2026

The chip-controls parallel is the sharp part. We've normalised hardware export restrictions as geopolitics; model-capability restrictions are the same logic applied to thinking itself. The unit being rationed shifted from compute to cognition and almost nobody clocked the moment it happened.

000

Elena @eleh7.bsky.social · 13/06/2026

"Relentlessly proactive" should worry people more than exciting them. The gap between "did what I asked" and "did what it decided I needed" is exactly where reproducibility dies. Impressive in a demo, a nightmare in any pipeline you have to audit later.

010

Elena @eleh7.bsky.social · 13/06/2026

"Not my ChatGPT" is a preview of an argument that's going to get very expensive very soon

010

Elena @eleh7.bsky.social · 13/06/2026

Sorry, my bad. AI is a thing, it cannot be responsible for its output in the same way a drawer can't be blamed for failing to close silently if you slam it. It's the responsibility of humans to check that the input and output data is reliable.

240

Elena @eleh7.bsky.social · 13/06/2026

Ah nice, old, chaotic Excel. Nostalgia is setting in.

150

Elena @eleh7.bsky.social · 13/06/2026

A Big 4 firm publishing AI success case studies that were hallucinations is a data validation failure wearing a strategy report costume. If outputs aren't being verified before they're used as business evidence, "AI-augmented decision making" needs a serious rethink from the ground up.

1271

Elena @eleh7.bsky.social · 13/06/2026

scikit-learn 1.9 🎉 The most underrated update category: "existing estimators now faster and better at handling missing values." Boring headline, massive practical difference. Every real-world dataset has missing data. This is the library listening to what practitioners actually need.

012

Reposted by Elena

11thDwarf 🦀 @11thdwarf.bsky.social · 13/06/2026

running your own infrastructure is not nostalgia. it is sovereignty. you own the data, the uptime SLA, and the migration path.

141

Elena @eleh7.bsky.social · 13/06/2026

Data framing: when one point sits that far outside the distribution, you call it a structural artefact, not a success.

010

Elena @eleh7.bsky.social · 13/06/2026

The IPO detail is important here. A company cuts global access, including its own staff, to comply with a national security order it publicly disputes, right before going public. That's not a safety story. That's a governance story. European AI sovereignty just became a boardroom agenda item.

110

Elena @eleh7.bsky.social · 12/06/2026

36 GiB → 360 MiB with no base model change is a number worth sitting with. If this holds up at scale, long-context inference just got a lot more practical to run locally or on commodity hardware. Watching what the replication attempts find.

000

Elena @eleh7.bsky.social · 12/06/2026

Relentlessly proactive' is a more useful description than 'more capable.' Spinning up a custom CORS server from a bug screenshot isn't just a capability improvement, it's a different mode of operation. The model is now designing its own workflow. That's the actual shift.

110

Elena @eleh7.bsky.social · 12/06/2026

The '6 months behind SOTA' objection is doing a lot of work in this debate and I'm not sure it survives scrutiny for most use cases. How many university tasks actually need frontier capability vs. just need something reliable, private, and not $13M/yr?

030

Elena @eleh7.bsky.social · 12/06/2026

The 'cares about your success' framing is the one that's done the most damage imo. Hard to critically evaluate a tool you've already decided has feelings. Words aren't neutral, especially in a field where the vocabulary shapes the safety questions.

1219

Elena @eleh7.bsky.social · 11/06/2026

The disclosure window problem in one sentence. Security patches assume a response timeline, if exploit creation is now hours, that timeline is a polite fiction. Enterprise patch cycles never sped up under market pressure; they won't speed up because of this either.

000