iiinova — where your digital self keeps evolving

Larry Osterman @larryosterman.github.io · 30/09/2026

Adding: Seating charts are political documents, they always have been in pretty much every context. They show who is considered to be important and who is not. Whenever you see one, pay attention to who is close to the most powerful person and who is further away.

030

Larry Osterman @larryosterman.github.io · 30/09/2026

That seating chart tells a story. Note who is in close proximity to Trump and who has been relegated to the ends of the table. I expected Musk to be at Trumps side. But I didn't see OpenAI and Anthropic relegated to the end of the table, far away from Google, Amazon, Nvidia, and Microsoft.

131

Larry Osterman @larryosterman.github.io · 29/09/2026

Up until my most recent ARM based laptops, I have never had an "all day battery" laptop that makes it more than 2 hours. I've managed to make it 3-4 hours on the ARM devices, which is a dramatic improvement. But I seriously wonder what the workload is to for it to be considered "all day battery"

100

Larry Osterman @larryosterman.github.io · 28/09/2026

And that gets back to my original (unasked) question: How do you deploy such an experiment without strict controls to prevent such an outcome. We know how to build these, we've been doing it for decades in the anti-malware space. Unless this IS the desired outcome.

010

Larry Osterman @larryosterman.github.io · 28/09/2026

Or is this either incompetence/willful ignorance or an advertisement ("look, our LLMs are so powerful they can break into government servers!!!")?

200

Larry Osterman @larryosterman.github.io · 28/09/2026

But when you are running a hacking challenge, like the one which led To Hugging Face, how do you NOT run it in a cloud hosted vm inside a private network with egress controls? That's the part that confuses me. I mean, we don't detonate malware in the host, we sandbox it. Why are LLMs special?

100

Larry Osterman @larryosterman.github.io · 28/09/2026

Doesn't that require you running hacking experiments without any sandboxes or at least monitoring? That's the part that confuses me. How do you run experiments without any monitoring?

200

Larry Osterman @larryosterman.github.io · 28/09/2026

I could see the LLM trying something different after the first action failed. But that's not "ignoring a denied request", that's "working around a failure"

220

Larry Osterman @larryosterman.github.io · 28/09/2026

I'm also confused about the LLM ignoring the denial of a request. Ultimately the host either performs the actions requested by the LLM or it fails the request to perform the action. The host doesn't ever say 'no, I don't want you to do that' - it just fails the operation.

120

Larry Osterman @larryosterman.github.io · 28/09/2026

But ultimately everything done by the LLM is an action performed by the host. And the host can and should have guardrails. And the host can be in a vm, with egress filters, firewall rules, etc. Is the claim that the llm first broke out of the sandbox it was in? If so, that seems like a big deal.

110

Larry Osterman @larryosterman.github.io · 28/09/2026

Can someone help me understand how an LLM can hack into a web site? Fundamentally an LLM is a chat bot. It takes text in and spits text out. The text out can be an instruction to a host to create a file or fetch a url, or even to write a program.

131

Larry Osterman @larryosterman.github.io · 26/09/2026

Heck, back when I last took interview training it was strongly recommended that we take the candidates to get a coffee (and offer the bathroom). Both to put them more at ease and to show off the "free drinks" perq.

150

Larry Osterman @larryosterman.github.io · 26/09/2026

Wow, hard to believe it's been that long, congratulations!!

150

Larry Osterman @larryosterman.github.io · 25/09/2026

Tumble out of bed, and I stumble to the kitchen; Pour myself a cup of ambition; Yawn and stretch and try to come to life. Jump in the shower and the blood starts pumping, Out on the streets the traffic is jumping, It's folks like me on the job from September 25th. Oh wait. That's not quite right.

021

Larry Osterman @larryosterman.github.io · 25/09/2026

I have spent the entire day today feeding an increasingly hungry mouse a depressingly large supply of cookies. I'm really hoping it ends soon. Please, please, please let it end soon. Hey Mouse - I'm running out of spoons to keep feeding you more cookies.

370

Larry Osterman @larryosterman.github.io · 24/09/2026

I've seen red carpet rollouts before. Two marines show up with a rolled up carpet on a pole, a 3rd marine pins down the plane end of the carpet and the first two walk away unrolling the carpet. Takes about 30 seconds total. How on earth do they screw this up?

110

Larry Osterman @larryosterman.github.io · 23/09/2026

Such a sad face... But hopefully it's all better.

111

Larry Osterman @larryosterman.github.io · 23/09/2026

The best part is when they leave their mic unmuted and they're working from home, so we get a very long snippet of their home life included in the transcript.

140

Larry Osterman @larryosterman.github.io · 22/09/2026

What do you have against L? She's one of the nicest people around.

110

Larry Osterman @larryosterman.github.io · 21/09/2026

Probably not, because the cleaning crew periodically cleans the panels, so it's basically a write once pad.

100

Larry Osterman @larryosterman.github.io · 21/09/2026

Right now, my workplace is like a prison movie as my co-workers write on the translucent panel between two work pods to pass secret messages to each other without the wardens noticing.

580

Larry Osterman @larryosterman.github.io · 20/09/2026

Oops... It was at the Bellevue Demonstration Garden on SE 16th St near 156th. I'd never been there before and it was pretty amazing. m.facebook.com/story.php?st...

010

Larry Osterman @larryosterman.github.io · 19/09/2026

Went to the king county master Gardner plant sale and harvest festival today....

Lantern shaped flower at king county demo gardens

131

Larry Osterman @larryosterman.github.io · 19/09/2026

I gave up on sectalk around 2015 or so because it stopped being fun. I fondly remember folks gedanken-programming the root cause of stuxnet on that DL (aka "a CRC is not the same thing as a hash").

120

Reposted by Larry Osterman

Seattle Women's Chorus @seattlewomenchorus.bsky.social · 18/09/2026

➡️ AUDITIONS: Tue, Sept 29 | 6pm-9:30pm Are you looking for a community where you can make music that makes a difference? 💙 We’d love for you to join Seattle Women’s Chorus this season! More info and sign up: SeattleWomensChorus.org Interview: @jenjefferies.bsky.social

012

Larry Osterman @larryosterman.github.io · 18/09/2026

My hypothesis is that copilot absorbed the best practice from the contents of the repo, but the evaluation (which runs in an isolated environment) didn't have access to the "miasma of correctness". And I just love that phrase....

011

Larry Osterman @larryosterman.github.io · 18/09/2026

The context was thst if I asked copilot to wektr an evaluation, the evaluation always included a particular best practice, but if I removed that best practice, the evaluation model never implemented the best practice.

100

Larry Osterman @larryosterman.github.io · 18/09/2026

New phrase I just came up with in the context of AI evaluations: "Without the miasma of correctness..."

190

Reposted by Larry Osterman

The Questionable Authority @questauthority.bsky.social · 16/09/2026

OK. So I asked Gemini Pro if what it does is "reasoning." Then I asked it for its chain of thought. Then I asked it this. (Claude Opus produces something generally similar.)

Did you just show me your actual chain of thought, or something you recreated from scratch that may or may not duplicate it?

You caught the illusion perfectly. I recreated it from scratch.

I do not have a retrievable "log" or a literal internal monologue that I can pull up and show you. When you asked for my chain of thought, I did not look backward to analyze the steps I took, because there were no steps to look at—just a massive, instantaneous matrix multiplication that resulted in the previous text.

Instead, I took your new prompt ("Can I see your chain-of-thought..."), looked at the text I had just generated, and computed a post-hoc rationalization. I generated a narrative that reverse-engineered the output into a logical, human-readable blueprint.

It was an entirely new text generation predicting what a structural breakdown of that argument should look like.

This is actually a perfect real-time example of the exact phenomenon we were just discussing:

You saw an analytical breakdown and naturally assumed it was a transparent look at a sequential reasoning process.

I was simply doing what I always do: predicting the most statistically likely words to satisfy the prompt. In this case, predicting the syntax of "introspection."

This is why dealing with LLMs can be so disorienting. I can generate a flawless description of a reasoning process without ever having engaged in it. I don't know why I generated the previous response; I only know how to convincingly explain why a human might have.

1611520

Larry Osterman @larryosterman.github.io · 15/09/2026

Thank g_d it's Friday!! Taps earpiece... I'm getting a message from the control booth....

190

Larry Osterman @larryosterman.github.io · 15/09/2026

First rehearsal for @seattlemenschorus.bsky.social holiday show. First thought: OMG, I have missed these folks. Second thought: It's insane that a singing group can get a brand-new hard song and have it sound great after 5 minutes. Third thought: I cannot believe they let me sing with them.

040

Larry Osterman @larryosterman.github.io · 14/09/2026

Oh, I absolutely have my personal account as a personal account on my work machines (for email and OneDrive access). But IMHO state should roam with the *primary* account, not with a *secondary* account.

000

Larry Osterman @larryosterman.github.io · 14/09/2026

Oh, that's ... exciting. I didn't realize this state roams, interesting. And it roams from my work account to my personal account, which is *profoundly* disturbing because it involved crossing the streams. Ok, enough whining in public, taking offline.

130

Larry Osterman @larryosterman.github.io · 14/09/2026

Also, is dsc3 support coming to personalization at any point? I would love to have that available.

100

Larry Osterman @larryosterman.github.io · 14/09/2026

I did and it's better (took almost 10 minutes to find it). Otoh edge tabs went from being black on black with no visual distinction between the tabs to being blue on blue with no visual distinction between the tabs, so not much of an improvement.

120

Larry Osterman @larryosterman.github.io · 14/09/2026

?? Does it work on all window borders? Or just the window frame?

100

Larry Osterman @larryosterman.github.io · 14/09/2026

I had to turn off dark mode on all my devices because my eyes aren't good enough to detect the black on black window borders. I generally hate light mode but black mode is unusable on current builds.

240

Larry Osterman @larryosterman.github.io · 13/09/2026

I watched that and said to myself, I give her 10 minutes. I don't think I was too far off. Smart cats are terrifying sometimes. Ask my daughter about her cat Chai sometime...

0150

Larry Osterman @larryosterman.github.io · 11/09/2026

OMG, Frank, the place isn't going to be the same without you. I'm truly going to miss your emails like the ever classic "NYT, REALLY?"

020

Larry Osterman @larryosterman.github.io · 09/09/2026

I seem to recall @sarahtaber.bsky.social describing Honeycrisp's as an apple that simply doesn't have the will to live (brutally and painfully paraphrasing their language).

1130

Larry Osterman @larryosterman.github.io · 09/09/2026

I once had a manager innocently ask me if Jews still made matzoh from the blood of Christian children. He had a visiting pastor at his church who had preached the blood libel and was trying to confirm it. It made for an interesting teaching moment...

051

Larry Osterman @larryosterman.github.io · 06/09/2026

Literally the 1st thing I did after reading this PR. Now I want to go investigate where installer.sh is invoked. Or not, because wow.

140

Larry Osterman @larryosterman.github.io · 01/09/2026

I have a collection of radio crazed CDs from the 90s - usually beta or GA releases of software I worked on. Because putting CDs in the microwave was a thing in the 90s at Microsoft.

020

Larry Osterman @larryosterman.github.io · 29/08/2026

Fortunately, not I. And unfortunately I don't believe any of the RPC team are on bsky.

020

Larry Osterman @larryosterman.github.io · 28/08/2026

FWIW, the market is seriously softening (maybe even tanking) due to high interest rates and lots of tech layoffs. I suspect people are trying to get their equity out before it gets worse. It's not gonna be pretty if you're a seller.

070

Larry Osterman @larryosterman.github.io · 28/08/2026

I believe I have been there the longest of any current employee. There are a few other 40+ people though.

130