ODSC Notes

I just finished a trip to attend the ODSC conference in Boston, leaving without really understanding where we are on the hype cycle of Agentic AI . I attended hoping to sample from the unknown like you usually do at big conferences, but there was no aha moment. I don’t think our industry is ready for an aha moment yet.

On the flight back, I read on article² where Rogé Karma was walking back a stance he had about the AI bubble given new revenue data he was observing. The conference gave me a lot of confidence that regardless of what benefit agent AI will ultimately have, there was now no doubt that companies and individuals who do not upskill will fall behind in one way or another. But the article put down some numbers about the new revenue that Anthropic, Open AI cursor and the data center companies they rely on, Microsoft, Google, Amazon, Core weave, were now recently experiencing, taking them perhaps, out of bubble territory.

I thought back to my notes from a talk that was challenging the big lab model narrative.

Private LLMs

One of the first talks I went to was about private LLMs, and during this talk the speaker from Datasaur.ai¹ said that the adoption of AI found a cheat code, agentic AI. He pointed out that, in crossing the chasm³⁹ terms, stats were showing only 15% of knowledge workers adopted chat organically, but when tools now added agents into the interfaces into themselves into the existing mediums people already use, like the ubiquitous Microsoft Copilot which is in the Windows taskbar, in all Office applications , in the Edge browser sidebar, then adoption magically went to 100%. This was the cheat code to side step organic conversion.

He called this the “opt out” vs “opt in” strategy. I asked, since his talk’s focus was about encroaching LLM cost, wouldn’t it be better to allow organic adoption to continue to play out instead? He countered that sometimes you don’t know what you’re missing and this shift is a useful nudge. In hindsight I would call this the 401k opt out trick many economists have recommended. I wonder what those economists would say about this LLM approach.

Perhaps unironically, as I was at my gate waiting for departure back home, I saw a post from a friend about https://copy.fail, the newly discovered Linux privilege escalation vulnerability, where a user mode process can su to root. So I keep this in mind while I learn about people excitedly giving anthropic access to rumage around with execute permission on their laptops.

Back to the speaker’s presentation, his point was that indeed the risk of opening up your company to Anthropic or other AI services, giving them a front row seat to your company’s data, can be avoided with the open weight models that are now becoming as good as their SaaS counterparts, and with that per token saas cost recently going up, along with surge pricing, also perhaps even a cost benefit too–Cory Doctorow’s AI enshitification³⁵ was falling into place. This is where you lock customers in so they find it hard to stop using your product, and then you reduce the value proposition, either by degrading the quality or raising the price.

Datasaur.ai¹ is not the only company discussing open weight models. In parallel, Cloudflare had Agents Week and released serving Kimi K2.6 on WorkersAI¹⁸.

The future of code

Between talks, I spoke with colleagues about the future of code. One was inspired by a conversation I had with a speaker on program static analysis¹¹, Armando Solar-Lezama, potentially making a come back. Or at least his research team was identifying that there was a gap in code evaluation, bug evaluation in particular. He noted that even Mythos–which he said he did not have access to–, would not find all the bugs, but for the sake of reliability, we do need to find all the bugs, or at least all the extra ones that code gen is contributing. So after his talk, I asked him what was his vision? I wondered, hey, python was so popular for data science, because it was so effortless to start learning, but also come with ease to create bugs that static type and memory safe languages like Golang do not run into. And it sounded like yes he sees there is definitely now more room for strongly typed memory safe languages like Rust, though that will still not be enough to make sure new code that is generated is reliable. I think one of my take-aways from his talk was that code generation is producing an unprecedented volume of code that needs to be reviewed and we desperately need better ways of vetting it for reliability. And hopefully code gen now frees us up to focus on bringing back precisely that kind of code analysis that was very popular in the early 2000s but then faded away.

One colleague I spoke with was in the audience for this talk as well . They had transitioned to our ml platform team from data science and, they found themselves in a coding skill plateau because they saw advances in code-gen made it pointless to grow their skills. This conference was especially telling them that skill is less in demand now.

Personally, I’m of the thought that most code I have generated, has had a low signal to slop ratio. Or at least it was too low for my slop tolerance. And so I have been so far finding³⁶, the main benefit in one off POCs or where I intentionally was building a non-production capability that I otherwise would not have had the time for. But I don’t see myself using code gen in place of real learning opportunities . And building without understanding what I’m building never came naturally to me. Or at least there’s a goldilocks sweet spot of understanding where I like to hang out. I get a similar impression from listening to Mario, author of Pi, per an interview with Gergley Orosz, he has said he doesn’t look at code generated, unless it is on the critical path.

Intentions chunked

Roge Karma points out⁴ in his article, citing a SemiAnalysis piece⁵, deconstructing knowledge work as chunks of Read, Think, Write and Verify. And that makes it a good candidate for building blocks in agentic flows, that can be learned, as long as the criteria are well defined. I pick things up I put things down. In other words, can knowledge work be cut up into units of work that are commoditized. I would flag here that this sounds remarkably similar to the vision of the waterfall software planning model that the agile manifesto⁶ of 2001 responded too, as well as the Data Science as Pin Factory article^7,8 from 2019 written in response to the desire to assembly-line-ify data science. The agile software movement pointed out that software projects are messy and customers cannot accurately describe what they want. And Eric Colson extended this to the messiness of extracting signal from data. In fact his description of the ideal data science pin factory echoes that SemiAnalysis article:

“one person sources the data, another models it, a third implements it, a fourth measures it”

During the conference, I was chatting with another colleague who was excitedly plotting how she can carve out some EDA time soon–Exploratory Data Analysis time–with an unstructured dataset she has been sitting on, using some new techniques the conference inspired her to try. She believes she would need at least a good 6 months, of, finding time in the cracks of her day job, to determine if there is enough there there in her dataset, before even proposing an improvement that her customer can consider.

In an interview⁹ with Peter Steinberger–creator of OpenClaw, an open source agent–, I listened to a few months back, he described his niche as “difficult but not too interesting”. This is precisely the opposite of low hanging fruit, the problems that are right there in front of you, easy to understand, easy to describe quick wins like the SemiAnalysis Read-Think-Write-Verify chunks. Or at least they are of much higher frequency and smaller scale than weeks or months. Peter responds to people who attempt to preplan a backlog of units of work, orchestrating a team of agents to coordinate and execute on the plan:

“I don’t believe this works. Like, this is the waterfall model of software building. This we learned long ago that this doesn’t work. Like, yes, people work differently and maybe it does work for some. I just don’t see how this could work for me. Like, I have to start with an idea and often I purposefully under-prompt the agent so it would do something that would give me new ideas. You like maybe like 80% of the things I assumed were like crap, but like there were like two things like, ‘oh, I didn’t think about that way.’

“And then I iterate and shape the project. And I have to click it. I have to, like, I have to feel it. I feel, to make good software, you know one thing those things often lack is taste. I have to feel like, how does this feature feel? And the beauty now is that features are so easy, I can just, like, throw it away or, like, re-prompt it. My building model is usually very much forward. It’s very rarely that I actually revert and have to go back. It’s just, like, ‘okay, no, then let’s change this. No, let’s do this.’ It’s like it’s like shaping. I love how this, like, you start with a rock and then you, like, chisel away at it and, like, pick different areas, and then slowly like this statue emerges out of out of marble. That’s how I see, that’s how I see building something.”

That is a reflection on the creative process. I think if anyone would, Steinberger would be a good judge of how agentic programming can massively speed up your experimentation loop, but it nevertheless is a loop you cannot reduce into a clearly defined deterministic sequence of units you can assign over to your army of agents.

Good bye project planning?

I hear Steinberger’s take, more than anything as, that agentic programming is the final nail in the coffin of using product planned roadmaps to derisk quarter long software development efforts.

Perhaps we can acknowledge though that there are still then two kinds of work in the themes of explore and exploit: spikes that are open ended that produce research artifacts and repetitive tasks that are more well defined because you have done them many times already. And agentic work can perhaps make the first kind easier to bound box. And the second kind, repetitive work, is a great candidate for code-gen, since the edge cases are much better understood.

An Internet for AI Agents

The first talk I attended was practical. It reminded me of a more fleshed out moltbook.com. Ramesh Raskar laid out a vision¹⁰ for how agents can communicate in the future. He pointed out that currently agents are clients and they do not have URI endpoints. And NANDA proposes a DNS for agents among other aspects.

He also made an intriguing prediction, that in the future, he expects every person will have multiple agents working on their behalf, in agent marketplaces. He had presented this vision for residents of India first, since in India, everyone now already has an Aadhaar card a new identity infrastructure, for a while now. And a proposal can be for agents to asynchronously in a decentralized way work on the ONDC and UPI. So this makes India very unique. Your agent can represent you in case people want to ask you questions and perform tasks on your behalf. But also he sees a kind of democratization. The agent is called a doot, which is a local word, meaning a messenger. This reminds me of the Shell Game season 1¹² actually, where the host attempted to give his agent access to everything about him and used more or less RAG and some agentic loops to handle various tasks for himself.

I think Ramesh’s vision assumes a lot, including the lynchpin that everyone will be an entrepreneur.

I like the part about freeing up ourselves to not deal with busy work. But yea I have my doubts about everyone managing their own agents. But intriguing nonetheless.

MCP Travel agency workshop

I also attended a workshop of mcp-toolbox, a Google library to get started faster building out your own MCP server. This session by Wenxin Du helped make a few concepts on building agent apps click for me.

A take away though is, you can build out skills for an agent to, say translate natural language to sql, which was the first iteration during the workshop of the travel agency example, but that was acknowledged as too risky, since there is a risk of someone running all kinds of unsafe queries on your database. So then through several iterations in the workshop, RBAC was added so agents powers are limited, and only to the access intended for the specific person who the agent is interacting with. And the natural language to SQL was also made more restricted, as well. So then you kind of end up with something that is quite similar to what we had before MCP, which was SQL prepared statements. So mcp-toolbox also uses its own flavor of parameterized restricted prepared statements³¹. So are we sort of back where we have started?

But then I did some more research on MCP

This field is clearly being heavily discussed this year. I was trying to track down one article on MCP’s demise I had lost the link to and found a lot of activity on this topic!

The interesting rift with the “MCP is dead” rift is that altbough MCP is a good protocol for an agent to discover what tools are avqilable in an environment and then to help route to them, and authenticate if and how a user is allowtd to use them , this is feeling like reinventing the wheel if this is a coding environment, because the tools are already pretty established for code. And on a local laptop its already just you and your permissions so authentication is not a mystery. And crucially, embeddings are not great at being exact with code boundaries , whereas ast was made for this. Embeddings can help find general areas relevant for code modification, but they struggle with fine grain edit points. shell tools like grep and find are more granulzr and then abstract syntax trees can help with precision too.

Kind of how a decade ago similarly you would say sure you could use keyword search to find and replace specific code, when using vim motions or sed, but you would make your edits more precise by using regex. That was the state of the art back then. Today, we have similar choices.

Back to why planning is hard

Should refer back to also Cal Newport multi step automation article¹³. Yea . Agent Harnesses and verifiability for each step.

Vibing Abstractions

Abstractions and modularity are a useful principle in software, helping to know the right level of information at any point in time. These days if you want to deploy a website, you don’t need to know that the internet is built on TCP/IP, OSPF/RIP/BGP, HTTP, SSL/TLS, DNS and the rest of the alphabet ocean of protocols.

What would Rich Hickey say here w.r.t. Simple Made Easy and locality of knowledge?

I would wager he would identify Vibe Coding as the ultimate Easy Button , leading to all kinds of under-the-hood complexities , which allow you to go fast initially, but then slow down dramatically as your tech debt accumulates.

slow is smooth smooth is fast

I suspect you can incorporate LLMs into your workflow , jumping up and down to the right level of abstraction as needed, without pretending to be Neo from The Matrix , downloading experience.

What do you need to know

Hearing also here²⁹ , similarly , “oh I learned DSA in school but rarely used it on the job”. And arguments about how, well just because it is not being used doesn’t mean it was not part of the journey, the learning journey. These are kind of the intangible effects I have heard a lot of other people go into as well.

I might have written this elsewhere already, but going super deep on whatever it is that you do instead of shallow, is perhaps more important than what it is . Because there are so many topics and so many technologies you start and change. But also people say there are those fundamentals. And skill transfer.

This connects to the no code topic too.

Large Code bases though?

Other than abstraction i realize another argument to not understand the code is obviously these days code bases arr super large anyway and no onr person understands it anyway. But the point is i think you can zoom in as needed and then learn.

And maybe another use of an LLM can be code understanding . At a moment notice. But the thing to understand again is you only have that ability, leverage your years of experience writing code. That’s when code can be FOIAble . You have to knos whzt questions to ask.

The enjoyment aspect

Can shallow work also can pick away at your soul or your ego? There will be many opinions on whether enjoying what one does is a form of entitlement and that some would say work is called work for a reason of course. However, if you do get a chance to go deep on what you do, it can lead to mastery and that is a big aspect of fulfillment. And technology has changed an endless number of times over the millenia. People have been developing their craft and constantly reinventing themselves over and over again. I come back here also to this book, Range, that is also on this topic.

The bitter lesson for software science?

But ML says hand wiring neural nets is a kind of artisanal joinery which will never outperform an algorithm like SGD . So is all software engineering just going to software science?

I recently had a conversation with a friend working at a company where he has gone full Claude Code a few months ago and hasn’t touched software. I forgot to ask him about the quality of the software, but his experience sounded like he was becoming a 4x engineer. Intriguingly, his work with Claude Code, did not speed up his flow on any particular project, but it introduced a chance for him to multiplex himself, context switching between up to four different projects. This sounded to me like Peter Steinberger in his early days of the agentic experience before he switched to rapid fire single tasking. My friend admited his 8 hour days now sometimes felt like 20 hour days, but he was giving it a 100% to understand the new style to its fullest.

AI Skill Flip

I was at the conference founders talk too, Sheamus McGovern, about his book AI Skill Flip. I think the flip is about your journey and open to interpretation and not necessarily my friend’s 4x engineer path.

Sheamus was like yea theres AI job washing, that there are Amazon layoffs he pointed to where you learn people remaining are picking up the slack. (And side note Amazon was in the news for AI code related outages. ) And yea its a hive mind. He was joking how everyone’s conference slides looked like they were generated by Claude.

But he also noted nevertheless you need to read the room. If your company is not doubling down and not helping you reinvent yourself , they will fall behind and take you with them. And reinventing is how he lived his whole life, flipping his career between tech and finance and then finance to data science and then investing. Pivot I guess I realize was the word perhaps he did not use but sounds right here.

And the Productivity Paradox

And Dawn Choo gave a talk about how you appear to be more productive with new tools and therefore you thought that gives you more breathing room. But instead you find more on your plate. She notes instead you are on the wrong part of the pyramid and should use the new AI space everyobe is uncertain about to gain more influence to do less work. Reading Her substack article on this topic too [25], reminds me the why. Because yes everyone is trying to catch up and especially knowledge gaps that leadership has at your company and you can help be the translation layer which data science has always played anyway as she underlines.

As well as the same non deterministic probabilistic output I hweard others point out too. Sort of a contradiction. That now you can dump out a loe of plumbing code but it can be subtly wrong so now you have to fix it. The “eval gap” she identifies.

Since I do see this bug introduction duscussion all throughout this conference I dont know why people arent giving the alternative, dont use this to generate your code necessarily since that is riddled with subtle bugs. Use it to speed up the reviews say instead . ( some companies like Code Rabbit do offer this actually ).

A jack of all trades?

So then what should you know? People talk about T shaped or TT shaped knowledge, where you have broad knowledge in many diverse realms and more specialized knowledge in one or two areas.

Does the Vibe era create an incentive to keep everything at a shallow level?

The ultimate generalist jack of all trades?

The other analogy is that of someone who leans into tech team management. In that role you shift your time from designing and executing on technical projects, to coordinating projects that a tech team is working on, tracking projects, finding and measuring gaps in execution, coordinating with adjacent teams and with executives. You can share your experience as an IC to level up your IC colleagues and you may try to keep your skills fresh from time to time, rolling up your sleeves, but ypur brain will proportionally emphasize your glue skills, and your execution muscle memory will become shallow.

To be fair of course anyone who plays an IC role , is still only hands on with a subset of technology and that subset also shifts as tech itself drifts.

So the open question then is code gen merely another such shift in kind. Just shifting execution to a different layer of abstraction, as with compilers say.

Or infrastructure as a service with terraform or aws cdk say. You now dont need to setup your own racks of servers and networks. You just provision compute . Maybe the answer is about compression. it it lossy or lossless?

Deterministic or non deterministic?

Makes me consider when corporations started to outsource work internationally in the 2010s. Teams spread across the US are with close time zones but across continents, the overlap is less and coordination tax goes up. But with Agentic outsourcing, there is no time zone difference.

Black box ? Explainable?

we have been using deep learning black box models a-la-the bitter lesson for a while now, agreeing an algorithm can create a better model than by hand tuned feature engineering. And we have accepted the black box nature therefore. As long as SHAPley can at least explain/interpret. Same for compiled code. How about code code? It is technically still readable. There is theoretically no reason why code code at least can be readable and minimal without being minified.

Precision

one of the things Armando said that also stuck, about stochastic nature of code. So hmm were producing all these fun bags of code but wouldn’t it be nice if we can be more preciss about vibing, yea sure spec driven development. Oh wait thats what code code was. Deterministic. Nice.

Mythos and the Halting Problem

On security now[16] , interesting comparison of Mythos to Y2K. However, host calls code as math, yet, my response would be that we have such a thing as the halting problem, where we know we cannot predict statically if a program will finish and so by extension, likely we cannot prove bugs are true or actionable statically (my hypothesis ).

Worse Than No Code

UML? No Code? That was crap. Here is no code take two.

Actually, vibe coded projects, have precursors, similar in incomprehensibility and or black-box-ness. That is, closed-source or DRM projects.

I believe Armando also made a perl joke during his talk as well; vibe code or code gen as write only code, is not only buggy but also high entropy, low signal to noise. Though verifiable. DRM, closed source, already has threatened open source , with Https://malus.sh , a la Evil Corp, potentially as a joke but maybe not.

But neural code next.js cloudflare vnext is already an example. [21] [22]. Though, to Mo Bitar point on nature vibe coding hands, with DNA being the neural code, and life being the unit test suite , yes but the robotics analogy is apt w.r.t. RL, since evolution has no simulation. And thats why it takes millions of years to add a opposable thumb. RL , interacting with the real worl, is the tricky part

A consultant. Create a passable demo then say bye.

Why is this relevant? Building greenfield is easy. Changing is hard. All the no code low code projects Ive worked on have this theme in common. The initial abstraction works during the demo. But then you want to add more data, change your data model, add another integration. All that is now gated behind your no code solution. Code code is powerful and infinitely extensible, and simply doesnt have this lock in problem. This is discussed in [29] too.

Hiring advice

The last talk I attended, by Arturo Natella [17] was about how most job postings cause the right people to self filter in their ambiguity. Lines like “Other duties” are read like “we havent figured out this role yet” . And that often there is a skill wall a “Big List” you are not actually interviewed against. This was less about AI than AI hiring because he was showing a few real job descriptions and then how he would change them. He called this “Key Performance Objectives” or KPOs, like an analog to KPIs.

The Tool Promise / Hope

Will we then just rely on tools getting better³³ like this 2025-12 moment Andrej Karpathy descrubes where suddenly Claude output has been less buggy for him? It will work soon it will be real good? He mentions we went from resolving counting r’s in strawberry with perhaps a hack (?) to other jagged edges aroun what code gen models with agentic harnesses had been tuned for with RL. And essentially his take is agentic engineering is taking the jagged results and making them high quality while benefiting from the speed/productivity boost.

Throughput thoughts

The log linear scaling laws were real. And I think prompt complexity to correctness throughput is likely also going to pop out as a real result. Riffing off Mo Bitar’s³⁴ note on a tweet result to this effect. And this seems to align with the pattern I was dicussing from Peter Steinberger’s⁹ experience too. He seemed to have landed on a kind of sweet spot for throughput of how much you specify, in a prompt, with the expectation of how correct the output will be.

A line from Mo around 7:15³⁴ or so was ringing true for me, that he would be uncomfortable shipping the code given its quality, well after actually looking at it, as opposed to just observing whether it passed tests. I suspect that feeling of retch is proportional to how close you feel to your customer. If you are in a very large company, you might feel more comfortable passing that code along but if you are in a smaller company, you have a higher sense of obligation to the quality.

More than RAG

Another practical talk I appreciated, was a session by Sara Zanzoterra on going from RAG to Agents.

A kind of foregone conclusion

The feeling I had leaving the conference wqs one of a foregone conclusion of the code gen path we were on. What no turning back now? This sentiment is echoed by Mo Bitar too⁴⁰. I don’t know what will happen, I don’t think anyone can really presume to know. But the conference at least felt less than neutral. Maybe was it a safe bet for them to make, perhaps by editorial choices of who spoke.

I’m glad I have spent some time to at least get other perspectives. I get the point about hedging, you should hedge and test drive the latest and greatest. But that also means hedge against abundance as well. As I’m writing this, the industry has now gone from several months of token maxxing to now getting more conservative. Tokens will get cheaper of course but pure abundance is an illusion anyway. Whatever you do, I tjink you need to do it deeply. Maybe multi task mastery of agents does become the new thing, that remains to be seen. Karpathy has dubbed it the end of vibe and start of agentic engineering right? Looking forward to seeing how it plays out, but without too many handicaps to my mind. It is a terrible thing to waste as they say .

References

1. Ivan @ datasaur.ai

2. https://www.theatlantic.com/app/webview/687022/

3. https://www.theatlantic.com/app/webview/687022/?app-privacy=apple-att&articleAudioPlaying=false&color-scheme=auto&fontScale=1#:~:text=As%20a%20group,the%20field%20unique.

4. https://newsletter.semianalysis.com/p/claude-code-is-the-inflection-point

5. agile manifesto

6. https://multithreaded.stitchfix.com/blog/2019/03/11/FullStackDS-Generalists/

7. https://hbr.org/2019/03/why-data-science-teams-need-generalists-not-specialists

8. https://youtu.be/8lF7HmQ_RgY&t=4180 , Peter Steinberger

9. Ramesh Raskar, https://nanda.media.mit.edu

10. Armando Solar-Lezama, “Open Challenges for the Next Generation of Programming Agents” , https://x.com/_odsc/status/2047815353967780118

11. Shell Game podcast, season 1

12. Cal Newport 2026 Feb New Yorker article about automation

13. Rich Hickey , “Simple Made Easy”

14. adversarial coding https://github.com/sepiariver/GAN-coding

15. security now, mythos, https://podcasts.apple.com/us/podcast/security-now-audio/id79016499

16. Arturo Natella https://www.goamaru.com

17. https://developers.cloudflare.com/changelog/post/2026-04-20-kimi-k2-6-workers-ai/

18. https://www.youtube.com/watch?v=Yxlb0-zTURo , RAmesh Raskar

19. Https://malus.sh

20. https://youtu.be/ateDMU5EGeg , Mo Bitar on next.js and cloudflare

21. Two Bubbles, the token costs and tokrn inflation Anthropic April 2026 https://open.substack.com/pub/techtrenches/p/the-ai-industrial-transformation?r=d7b46&utm_medium=ios

22. Dawn Choo , https://open.substack.com/pub/askdatadawn/p/the-data-career-is-evolving?r=d7b46&utm_medium=ios

23. OpenCode founder on productivity with AI code tools, https://blog.codacy.com/the-creator-of-opencode-thinks-youre-fooling-yourself-about-ai-productivity

24. interesting augmentation radiology example, 11:03 , https://youtu.be/eSABedBwZjQ

25. https://uxplanet.org/mcp-is-dead-cf16b667ba6d

26. “I don’t like Programming” | Prime Reacts, https://youtube.com/watch?v=r6EXZcTJyaA

27. Book, Range

28. https://mcp-toolbox.dev/documentation/configuration/tools/

29. The one shotting interview idea, Mo Bitar on Andrej , https://youtu.be/ZugX7a99dLk?t=308

30. Andrej, ok, https://youtu.be/96jN2OCOfLs

31. https://youtu.be/SKTsNV41DYg

32. Cory Doctorow, https://youtu.be/r_ktaPutkjM

33. benefits , http://michal.piekarczyk.xyz/note/2026-05-14--mine-the-gaps/

34. constraints, not just tests, https://www.augmentcode.com/guides/harness-engineering-ai-coding-agents

35. agentic coding with Pi, but slow the eff down https://www.youtube.com/watch?v=RjfbvDXpFls

36. https://en.wikipedia.org/wiki/Crossing_the_Chasm

37. https://youtu.be/-cc9OsZLKFo?t=165 , Mo Bitar, on the presumed choice of the trajectory of code gen. Like a foregone conclusion

^x