Larry Osterman

Larry Osterman

@larryosterman.github.io

Long term MSFT employee, Dad, Singer. Expect cat pictures, tech nerding, etc. Previous: SMB filesystem, Exchange Store, Windows Audio, Windows Runtime, COM, Confidential computing. Current: Azure SDK for C++ Proud LGBTQIA+ ally.

2.4K followers564 following1.6K posts

You know it's a good day when you can start a message to a coworker with: "Hi <coworker>, I know that you thought you'd finally gotten rid of me when you moved to your new job, but like a bad penny, I keep coming back." :).

Larry Osterman reposted

Doctors and nurses are voicing their objections. I-638 would require EVERY girl who wants to participate in sports to go through biological sex verification. That means expensive, invasive, and utterly gross examinations. Protect our kids and let them play. Vote NO on I-638

rangemedia.coMeasure to require "biological sex" verification for all WA girls’ sports athletes on November ballotIf passed, the hedgefund-millionaire-backed ballot measure would enact one of the nation’s strictest examination requirements for girls’ athletics.

The mcp team is under my manager and for a month or so I worked on the team before being moved to a different effort. So I have some commits in this release 😊

Larry Osterman reposted

An aside: these "shorthand" labels for long words like "accessibility" or "internationalization" are called "numeronyms." They work like this: take the first and last letters and put the number of omitted letters in the middle. Thus, "localization" becomes L10N.

Asa Dotzler@asadotzler.com

Some of the hardest things to do well in software are localization (L10N), accessibility (A11Y), and internationalization (I18N). Getting these right, however, pays huge dividends in usability and reach. Integrate them into your planning, research, design, and, of course, development and testing.

Larry Osterman@larryosterman.github.io

Oh, and remind me of the time a team I was on used cryptprotectdata to encrypt a secret on a vm managed by vmss. Once a month (reliably), the service failed because the key was corrupted. It took us WEEKS to figure it out.

OK, so what was happening was that when VMSS handled updates, they decided that rather than run the normal windows update process (which takes the machine offline for some amount of time), instead, they took a clean new vm and copied the user data onto the new vm.

Which worked incredibly well, but the machine specific crypto keys used by CryptProtectData changed and any data protected with CryptProtectData was now unreadable... And now we know.

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.

View full thread

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.

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.

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.

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.

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.

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.

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....

Larry Osterman reposted

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.)

Will Stancil@whstancil.bsky.social

Again I would highly recommend you use a modern model for something kind of open-ended and then watch its chain-of-thought summary

Larry Osterman reposted

💖 In honor of Dolly Parton’s extraordinary life and legacy, Seattle Men’s Chorus is re-releasing our full filmed concert, "Dolly," for a limited time. All proceeds will be donated by Seattle Men’s Chorus to Dolly Parton’s Imagination Library. Get a streaming pass here: bit.ly/4geLoKA

Sometimes I wonder how many engineering system improvements are driven primarily by anger... As in "This @#$@#$@# release script required me to do all this extra work, F-ck that s-it, I'm rewriting it".

If you live in the Seattle area, come sing with us!

Seattle Men's Chorus@seattlemenschorus.bsky.social

Being part of SMC is about more than singing in harmony with 200+ voices. It’s about finding your people, making connections, and being part of a community that loves to sing. 💖🎶 Auditions are Wed, Sept 9, from 6:00–9:30 p.m. and Sat, Sept 12, from 9–11 a.m. Sign up: SeattleMensChorus.org

Sigh. I just had to explain to chatgpt that it doesn't have to add asserts to confirm that the compiler generated correct code. And it *insisted* that it was in the right to add: if (foo is not null) { assert(foo is not null); : }

Two quick stories, both from essentially the same time (early 1990s). #1 The Windows 3.x team was told by their leadership that they needed to use the lowest end machines available to do their development because if they had low end machines they would make sure their product worked great on them.

View full thread

To be clear: Both viewpoints have significant merit, especially in the 1990s when Moore's Law ran supreme. Machines WERE getting faster and faster and RAM costs were dropping so provisioning a machine that could run NT 3.1 was not expensive.

And obsessing over performance and forcing engineers to use the kind of machines that existing customers use can be a powerful motivator to ensure real-world performance. So I kind-of agree with both points of view.

Two quick stories, both from essentially the same time (early 1990s). #1 The Windows 3.x team was told by their leadership that they needed to use the lowest end machines available to do their development because if they had low end machines they would make sure their product worked great on them.

#2 At the Windows NT 3.1 ship party, the lead for Windows was asked about the fact that NT 3.1 required much more memory than Windows 3.x and replied basically that "That's ok, memory prices will drop so it won't be a problem"

What's maybe worse is that it wasn't until I was in my mid-50's that I learned the second half of "how sharper than a serpent's tooth". My father used to say the first part when I (or my siblings) were being snippy. And I didn't discover the 2nd half until it was too late to compliment him on it.

Larry Osterman@larryosterman.github.io

I got to use Banquo's Ghost in an analogy at work and was saddened to realize how few of my coworkers understood the reference. At least they get it when I use "how sharper than a serpent's tooth"....

I got to use Banquo's Ghost in an analogy at work and was saddened to realize how few of my coworkers understood the reference. At least they get it when I use "how sharper than a serpent's tooth"....