Ben Labaschin

Agents in Production 2025 - Event | MLOps Community AI Agents Are Already Working — Let’s Talk About ItAgents are no longer just experiments. From e-commerce to customer support to analytics workflows, they’re quietly getting real work done in producti...

Fixed registration link: home.mlops.community/public/event...

Register:
home.mlops.community/public/event...

If you missed the first one, check out What Are AI Agents: learning.oreilly.com/library/view...

Writing another O'Reilly report! 📚 This time: Managing Memory for AI Agents

Giving a talk about it at @mlopscommunity.bsky.social on July 17 @ 1:55 PM

Memory management is becoming critical as we build more sophisticated agents. Will dive into challenges & solutions.

Figure 2 from paper "Agenda Seeding" shows a decision-tree like diagram in protesters choose nonviolent or violent tactics and the state responds with violent or more restrained forms of force. Figure 2 summarizes the model. In the first period, subordinate group activists attempt to raise awareness through agenda seeding that involves activists employing tactics, possibly nonviolent or violent, which might make the news. News organization still set public agendas but social movements, like any group seeking greater attention or publicity, attempt to entice media to prioritize their concerns. Southern civil rights protests, for example, were often planned for mornings so footage could be flown to New York in time to air nationally on the evening news (Torres Reference Torres2018). Similarly, movement leaders selected Selma, AL, as a site for voting rights protests partly for its proximity to television network affiliates in Montgomery (Torres Reference Torres2018). Media attention, however, can be sympathetic or hostile. Even presented with similar facts, different reporters and news organizations often structure stories in systematically different ways (Davenport and Litras Reference Davenport and Litras2003). Civil disobedience, for example, can be covered as a form of justifiable protest or unjustifiable criminality. Kinder and Kam (Reference Kinder and Kam2010) find that whether an issue can overcome societal biases to become salient depends significantly “on the ability of the issue to command the public’s limited and fickle attention and on how the particular issue is framed” (38). Activists, attuned to the ways in which media can frame an issue, routinely fought to craft narratives through a kind of political theater in which protests were “staged” and injustice “dramatized” (King Jr. Reference King1964; Lewis Reference Lewis2017).

Now, imagine a simple model in which activists opt to engage in nonviolent or violent resistance and then the state responds with typical or excess force. (To be clear, the world is more complicated, but these simplifying assumptions help us describe some common patterns.) 3/

Thanks to @danielrock.bsky.social , Tyna Eloundou, Sam Manning, and Pamela Mishkin!

I also wrote a quick write-up on where this research may lead us and the implications for labor inequalities here: econoben.dev/posts/extend...

An interesting finding: big gaps between our exposure measures and what companies report for adoption, suggesting much AI use is happening organically at the worker level.

We took occupation-level AI exposure data and mapped it to the actual workforce composition of ~8,000 publicly traded companies. We found the average firm has about 17% of tasks exposed to LLMs alone, jumping to 47% when you factor in partial integration with tools like Copilot.

I'm excited to share our new firm-level research on how AI exposure varies across US firms, published in AEA Papers & Proceedings.

My co-author Sam Manning wrote a great recap of our findings here: x.com/sj_manning/s...

Paper is out in AEA Papers & Proceedings: www.aeaweb.org/articles?id=...

If we haven't heard from JD in a month, we can safely assume he's somewhere in the Marshall Islands doing something cursed with monte carlo simulations and de-anonymized data.

Uber Eats really wants me to go to Applebees.

Oh damn. I don’t know why I didn’t consider that…

Well, at least there’s a Taco Stand nearby. Get a burrito.

Didn’t realize my man’s in SD.

Jury would be lucky to have you. I just got passed over for my recent selection bid.

Was on vacation with family for a week. So naturally, I built a new personal website. With all the chaos out there, I figured it was nice to know that SWE's are still out there rebuilding their websites for the thousandth time.

econoben.dev

Get off the Metra at Ogivile, step into Chicago. Guy across the street, standing by himself, holding a giant, bloody sign. yelling at people for their sins.

guy rides by him on a divvy. shouts “kill yourself!” keeps riding.

I think to myself: “ahhh, I’m home."

If people start a sentence with “Bill Ackman says…” to justify something ever again I’ll just share this tweet.

Have you done awful shit to English? And, if so, why not do so more often? 😏

Hey, journalists, here's an idea: when an elected representative claims to want to "test" the constitution, why not...follow up on that?

Here's an example:

BRENNAN: We should test the constitution.

GOOD JOURNALIST: Hold on. You're telling the American people you want to *test* the constitution?

trying out o3-mini-high. If you think this inner monologue is rough, you should see mine.

Wario. Would you say that it was gone in a SNAP.🫰 ??? … Wario, where are you going? Would you say it was gone in a snap Wario?! 😭 😭 😭

a close up of thanos ' face in avengers infinity war . ALT: a close up of thanos ' face in avengers infinity war .

Vic. Vic! Would you say it was gone in a SNAP. 😭 Vic, would you say that???

I choose a random word every time. Yes there are better words to start with but I do okay.

Yall probably out here thinking “squid” is a bad choice.

Starting my year off by demonstrating how to serve your own, private LLM you can access from anywhere in the world. A nice little engineering challenge I undertook while on break.

If you have any questions, feel free to leave a comment.

benjaminlabaschin.com/host-your-ow...

I think this law is limited. Consider an LLM with reinforcement learning whose reward function is not being found out as AGI. Such a system would be AGI without being determined as such.

Done better, done worse. Username: econoben

‘uv pip install torch’

Next. 🦦

I've been curating some of these "algorithms" since 1st grade if you really must know.

some of y'all who work in recommendations still don't get it.

the key to recommendations is to get your friends to text you videos and memes instead. It's called a "curated third-party algorithm" 💅