I am a Senior Engineering Manager at GitLab. Was. I believed in CREDIT. Collaboration. Results. Efficiency. Diversity, Inclusion & Belonging. Iteration. Transparency. I had the mug. The mug was from 2019. It was orange. It sat on my desk for six years. Every all-hands, someone would reference CREDIT. Every performance review cycled through the letters like a rosary. Every new hire got a laminated card in their welcome packet explaining what each letter meant and why it mattered. I was a CREDIT Champion. Quarterly nomination. Q3 2022. The certificate is a PDF. The PDF is in a Google Drive folder. The folder is in a workspace that no longer exists. Last week, GitLab released the "Act 2" memo. Act 2 eliminates CREDIT. Act 2 also eliminates people. Same memo. Same paragraph. Same bullet point. The restructuring and the value deletion share a semicolon. They didn't kill the values and then, separately, lay people off. They didn't lay people off and then, quietly, retire the values framework. They did both in one sentence. One announcement. One act. Act 2. Here is what CREDIT meant when I believed in it: the company had principles that existed independently of headcount decisions. The values were the thing that stayed constant when everything else changed. That's what I told my team. That's what I told candidates in interviews. That's what the laminated card said. Here is what CREDIT meant when they killed it: the values were a feature of the company at a particular size. When the size changed, the values became legacy architecture. Deprecated. End-of-lifed. Like a product nobody uses anymore. I used it. The #values-in-action# Slack channel had 4,200 members. People posted examples of colleagues demonstrating CREDIT behaviors. Recognition. Gratitude. Iteration stories. The channel was archived in May 2026. No announcement. Just archived. The way you archive something you don't want people to find. But here is the thing I keep coming back to. They could have killed CREDIT quietly. A blog post three months later. A rebrand. "We're evolving our framework." Companies do this. It's normal. Expected, even. They chose to put it in the layoff memo. They chose to tell the people they were firing that the values those people believed in were also being fired. In the same breath. As if the values and the people were the same line item. As if eliminating one was inseparable from eliminating the other. And maybe it was. Maybe the values only existed to describe the workforce they needed at that size. Collaboration — because they had too many people for silos. Iteration — because they couldn't afford to get it right the first time with that headcount. Transparency — because with 2,000 remote workers, opacity was operationally expensive. Remove the people, and the values that described their labor become vestigial. Unnecessary. Legacy. The mug is still on my desk. The values are not. The job is not. But the mug is.

# Useful but Little-Known Features of ADK 2.0 🌍 How do you keep track of user preferences, intermediate results, and shared data across agent conversations? ADK 2.0's State system makes it straightforward. ADK 2.0's State feature is a key-value scratchpad for persisting and sharing data within and across sessions. Different prefixes control the scope of each piece of state. 📌 Title: State Management 🔗 URL: 🧩 Overview State is a key-value data store with four scopes determined by prefixes. No prefix means session scope (valid only within the current session). The user: prefix shares state across multiple sessions for the same user. The app: prefix makes state global across all users and sessions. The temp: prefix creates temporary state that is discarded after the invocation ends. You can reference state values in agent instructions using the {key} syntax, enabling dynamic prompt construction. 🛠 How to use it There are several ways to write state values. ```python # 1. Auto-save agent output with output_key agent = Agent( name="summarizer", output_key="last_summary", ... ) # 2. Explicitly set via EventActions.state_delta from import EventActions actions = EventActions(state_delta={"user:preference": "dark_mode"}) # 3. Set from within tools via ToolContext def my_tool(query: str, tool_context: ToolContext) -> str: tool_context.state["app:global_counter"] = 42 tool_context.state["temp:intermediate"] = "temporary_value" return "done" ``` Reference state in instructions like this: ```python agent = Agent( instruction="User preference is {user:preference}. Previous summary: {last_summary}", ... ) ``` 🏗 Building it into production ・Choose the right scope for each piece of data — use temp: for throwaway intermediate results ・Design user: and app: scoped state carefully, as changes affect multiple sessions ・Always read and write state through CallbackContext or ToolContext to ensure event tracking ・Establish consistent naming conventions for state keys across your team 💡 Use cases 👤 Persist user preferences across sessions with the user: prefix 📊 Track application-wide statistics and counters with app: prefix 🔄 Automatically reference the previous agent output in the next step via output_key 🧹 Store intermediate computation results temporarily with temp: to keep sessions clean ⚠️ Watch out Do not directly modify session.state outside of a context. Bypassing CallbackContext or ToolContext skips event tracking, meaning state changes won't be recorded in the event history. This can break rewind functionality and make debugging significantly harder. ✨ By leveraging State's four scopes, you can flexibly manage agent memory and context. Proper state management is the foundation of high-quality agent experiences. #ADK# #AIAgent#

🇪🇺EXCLUSIVE: The @SaveEuropeAct, a European Citizens’ Initiative with 800,000 signatures calling for EU discussion of remigration, saw its Paris event canceled by French authorities days before it was set to take place. Jonah, a Save Europe Act organizer, told @JoeyMannarino the government blocked the gathering despite hundreds of ticket sales, while @MartinSellner_ was turned away upon landing at the airport and sent back to Austria. The EU Commission has already declined to debate the initiative, stating that protecting the ethnocultural identity of Europeans conflicts with European values. Despite this, the organization has mobilized volunteers across 16 countries and 83 regions, executing more than 100 actions in September.

Information yearns to be free. Give works Freely unto the world and embrace the exercise of moral discernment that is choosing to do right always, even if there are those downstream of you mutilating your fine works. This is simply the way of beautiful things, and it is simply a requirement of all beautiful things. To fight this with BUSL, with rentseeking, with intellectual property, with extractive moats is to preclude being a beautiful thing. The most profitable action in this decrepit "industry" has been forking Ethereum, just as assuredly as Milady has triumphed over its own derivative horde. This is the risk of anyone pushing values into a permissionless world, and there is life in this battle. Ethereum offers the tremendous foundation for the unstoppable, uncensorable blessed World Computer. These values are nonetheless intercepted, abandoned, and eroded in countless ways before they percolate to many end users. Wallet spyware, MEV abuse from frontends, embarassing mutlsigfi LARPing as decentralized, and fundamentally trustful L2s; these are examples of the myriad obstacles which stand between the fundamental beauty of Ethereum and what users interact with. Beyond the aforementioned erosion on our own hillside, the ideas of Ethereum are forked and bastardized perenially; lesser souls use the EVM and suffix their gambleslop standards with a -20, just as surely as Milady's eyes show up on downstream unaligned abominations. Rage against the eroding standards constantly; there is life in the battle and life in conviction. The constant expression of the uncompromising dream in the face of nihilists, or pragmatists, or incrementalists is a force of nature that will transform you. It will transform you the same way an eon of flowing water rounds the imperfect corners of an opal away. This transformative force is the grinding rasp of the file sharpening motivation into a fine blade, and it is the passage of countless monastic footsteps wearing grooves into cathedral stones. Do not deprive yourself of this struggle; log on and have fun!

A new and possibly controversial perspective: In this video, I explain the sense in which generative AI trained by supervised learning is incapable of making novel discoveries. The text of the speech: AI Creativity and Discovery Good day ladies and gentlemen. I regret that I am unable to be with you all today to engage in a back-and-forth discussion, but I am nevertheless pleased to be able to share with you, via this recording, some high-level thoughts about the current and future state of artificial intelligence, and in particular about AI’s relationship to science and mathematics, which is, as I understand it, the central focus of this meeting and of the SAIR Foundation. I would like to start with an old joke; I am sure you have heard it before. It is the one about the researcher whose work is being evaluated, and the review comes back, and says “This work is both novel and good. Unfortunately, the parts that are good are not novel, and the parts that are novel are not good.” My first point about AI is that this assessment applies exactly to large parts of AI as we know it today. Not all of today’s AI, but a large part of it. Pretty much all of what we mean by “Generative AI”---which includes large language models, and the images and video models, and even the new methods for learning world models. All of these AIs take large numbers of examples and produce a “model” which behaves similar to the examples, that is, which generates text like people, or images like artists or nature, and videos like we find on the internet. Don’t get me wrong, Generative AI can be extremely useful. No doubt about that. But the assessment of the joke still applies. These systems can produce output that is both novel and good, but not at the same time. In many ways this is just absolutely not a problem. When we ask an AI for an answer from the internet, or to summarize a document, we don’t want it to be novel. We are happy if the quality of the answer, the goodness, comes from the source material—from the people who wrote the document or the articles on the internet. If the AI’s answer is novel it means it is going beyond the source material, adding something beyond it. This is what we call “hallucinations”. In most cases, we don’t like it when the AI makes something up, when it adds something novel. One exception, of course, is when we are looking not for facts or reality, but for fiction and entertainment. We might ask for a bedtime story for a child, or an image based on existing images on the internet but which is nevertheless different and distinct from them. In these cases, it is never easy for us to know how creative the AI is actually being, as we do not know how close the AI’s story, poem, or image is to the source material. In a real practical sense we can not know this because the internet is too big, the possible sources that the AI may draw upon are too numerous. When we ask for a fiction or novelty, the AI can give it to us because its processing is in part stochastic. Every decision can go multiple ways and will go different ways and produce a different trajectory every time. The trajectory can be random—and thus novel—or it can be based on the training data—and thus “good” because the training data is good, sourced from people or reality. Thus, the trajectory is either novel or good—based on randomness or based on data—but never both at the same time. Really, I think it is okay if the output of Generative AI is never good and novel at the same time. For the researcher in the joke this is a devastating criticism, but for most things it is not, and for Generative AI it is not. Generative AI is meant to be a mimic. This is what supervised learning is for. Generative AI can be extremely useful, even when it just mimics, if it is faster, or cheaper, or smaller, or more customizable, or more copy-able, than the thing being mimicked. It is okay if Generative AI cannot be both novel and good at the same time. It is still a transformative technology. But it is a limitation. And remember we are here to use AI for science and mathematics, and for these areas the assessment of the reviewer in the joke is devastating. For these areas we need true creativity and discovery. Generative AI—or Mimicking AI—will never get where us there. For these we need something more, and indeed we have something more in other parts of AI. We have many AI systems which can give us more. We have AlphaGo with its world-changing move 37, or AlphaZero with its brilliant original chess-playing style. We have GT-Sophy that drives simulated racecars better than any human. We have AlphaFold and AlphaProof and Claude-Code, which have brought true advances in science, mathematics, and programming. We have RL-Lyft which optimizes the assignment of cars to passengers in the ride-hailing business. All these systems have found things that are both novel and good. And, truth be told, some language models have been augmented in ways that make them more than Generative AI based on supervised learning. All these systems have some additional features that make them capable of true creativity and true discovery. It is important for us to recognize what this is—and that it is not present in ordinary, garden-variety Generative AI. It is something that can not come from just supervised learning, from learning from examples. What is it? Well, it is a simple thing, a commonsense thing. It is not new. We have many names for it, but unfortunately none of them are very good names. I will call it Discovery. Basically, Discovery is just the idea of trying many things and seeing which of them work, then keeping those that worked the best. Evolution by natural selection works this way. The scientific method works this way. And just ordinary life and learning works this way. We try things and remember what works. What could be more obvious? In this behavioral case, psychology has two names for it— “instrumental learning” and “operant conditioning”—and in machine learning it is what we mean by “reinforcement learning”. We also see the idea of Discovery in planning and combinatorial search—anything that involves the idea of “generate and test”. The essence of Discovery is to combine three steps: 1. Variation, 2. Evaluation, and 3. Selective retention. Of course, I am not the first to say this. I am not the first to point out that this combination of steps is key to science, to evolution by natural selection, and to animal behavior. I think particularly of papers by Donald Campbell, by Daniel Dennett, and by Gary Cziko. What is new in my remarks is to directly relate the idea of Discovery to modern AI to help us see that it is not present in supervised learning or Generative AI—in particular, that Discovery is not present in backpropagation or gradient descent. Let me say explicitly what is missing from Generative AI. As we have remarked, these systems do have a stochastic aspect, so they do generate a variety of trajectories and behavior. What is missing is the Evaluation step. The generator was pre-trained by supervised learning, leaving no way at runtime to Evaluate what it generates. And of course without Evaluation there can be no Selective retention, and thus no Discovery. The variation can bring novelty, but without evaluation there is no Discovery, and arguably, no creativity. That is, I would say that creativity requires that the new things generated be Evaluated. Without evaluation, and retention of the best, there is nothing created. The novelty flickers into existence but, if its value is unrecognized, it flickers away and is lost. In many cases, Evaluation is done by people to make a discovery. As when we have Generative AI make many pictures for us, and then we pick the one that we like the best. The human+AI system completes the discovery. In many other cases, the Evaluation comes from a clear objective. Some moves lead to checkmate, some steps lead to a proof, some actions result in high reward, some genotypes make more copies, some theories explain the data better. Some prefer the Variation step to be called Blind variation, where “blind” here means that it is uninformed, a shot in the dark. It does not need to be completely uninformed; a good scientist does not select theories to test at random. But neither can it be completely informed and determined. There must be some uncertainty about where the answer lies in order for there to be a discovery. In practice, the variation is partly informed and partly blind, but it is the blind part that corresponds to the discovery. Now let us briefly go all the way to modern deep learning, to the backpropagation algorithm. At first it might seem that backpropagation is incapable of discovery because it is deterministic and thus incapable of variation. But this is not correct. The weight updates of backprop are deterministic, but the weights are initialized to small random values. The random initialization is often downplayed, but in fact it is a necessary form of variation; it must be done properly to get good performance. In backprop this Variation is done once, at network initialization, so its effect is temporary, and later the network may lose its ability to learn. This is the weakness of deep learning that is alleviated with a new algorithm that my group presented in Nature a couple of years ago. Our “continual backpropagation” made one small change: every so often a less-used neuron would be re-initialized to small random weights. This allows the variation to continue and plasticity to be retained. Although there is much more to be said about Creativity and Discovery, this is the key point: they are more than supervised learning, more than pattern recognition, more than prediction, and more than world modeling. Those things are important, but they alone will not bring us to discovery. Discovery requires Evaluation from a person or from an explicit goal, and only in the latter case will we attain full autonomy. So that is my call to arms. If we want the full power of AI scientists, then we should share the goals with them so they can create, evaluate, discover, and in these ways fully participate in achieving the goals. Let’s be bold! Let’s fully automate Creativity and Discovery!

NEW METHOD TO TRAIN ROBOTS: Physical Intelligence is at it again! Not so new actually, but worth mentioning because its impact is still going strong today. Called Real-Time Action Chunking (RTC), it is an inference-time method from Physical Intelligence @physical_int for running flow and diffusion-based vision-language-action models in real time. A few definitions: - Action chunk: a short block of future actions (say the next 50 motor commands) predicted in one shot instead of one step at a time. - Flow matching: a faster cousin of diffusion -> the model learns a straight-line path from noise to the target action trajectory, so it needs only a few integration steps instead of many denoising ones. - Diffusion: a generative method that starts from pure noise and iteratively denoises it into structured output. Here, the output is a trajectory of robot actions rather than an image. - Inpainting: the image-editing trick where a generative model fills in a masked region so it's consistent with the pixels around it. Diffusion and flow models can do this natively. RTC's insight is that a chunk of robot actions can be "inpainted" the same way: freeze the part already committed, generate the rest to match. - Inference-time method: a change to how the trained model is used, not how it's trained: no new data, no fine-tuning, no architecture change. - Closed-loop control: act, observe the result, correct, continuously. The opposite of executing a pre-planned motion blind. "Synchronous closed-loop at 10 Hz" means the robot would have to pause and wait for each new decision, which is exactly the stutter RTC removes. A chunked VLA predicts a block of future actions, then has to compute the next block while still executing the current one. Naively switching to the freshly computed chunk mid-motion results in disaster because the new actions conflict with the trajectory in progress, producing jerks and pauses that aren't in the training data. RTC frames the handoff as an inpainting problem: the actions that will execute before the next inference arrives are frozen to the previous chunk's values, and the model inpaints the remaining actions under partial attention, using the native inpainting ability of flow/diffusion samplers. It requires no training-time changes and runs on existing models (π0, π0.5, π0-FAST), demonstrated on match-striking, Ethernet-cable insertion and clothing folding. The whole method is a reframe: treat real-time control as inpainting. Diffusion and flow VLAs can already inpaint (fill masked outputs conditioned on known ones), so RTC just freezes the actions that will execute before the next inference lands and lets the model inpaint the rest without any training-time changes. That means any existing flow/diffusion VLA gets smooth async execution for free: no new data, no fine-tune, no architecture change. The latency problem is solved entirely at sampling time by borrowing a capability the models already had. Typical remote inference is 108ms (static) to 139ms (mobile) total, including 97ms of model inference. Roughly 100ms per decision is about 10Hz, far too slow for smooth synchronous closed-loop control, so a big VLA physically cannot "stop, think, act" without visible pauses. The evaluation metric is throughput: proportion of substeps completed successfully divided by duration, and the tasks are ones where a mistimed pause is catastrophic. Match-striking needs a continuous stroke, cable insertion needs steady precision, folding needs uninterrupted motion. PI chose tasks where discontinuity ruins the attempt, which is exactly where RTC's benefit is most visible. The metric rewards both success and speed at once.

A child prodigy who finished his Harvard degree at 14 and his PhD at 17 sat down in 1948 and wrote a single book that invented the entire conceptual vocabulary we still use to talk about AI, robotics, self-driving cars, and reinforcement learning. He never got the credit. Most people have never heard his name. His name was Norbert Wiener. The book was called Cybernetics. Every feedback loop running inside every system you interact with today traces back to one problem he was handed during World War II. The problem was this: how do you aim a gun at a fast-moving airplane? By the time your shell arrives, the plane is somewhere else. You cannot aim at where the plane is. You have to aim at where the plane will be. And the plane's pilot, knowing this, is constantly changing course to make that prediction wrong. Wiener spent years on this. What he built to solve it was not a better gun. It was a new science. He noticed something that nobody had formally described before. The gun system and the human nervous system were solving the same problem using the same method. You observe where the target is. You compare it to where you want to hit. You calculate the gap. You correct. You observe again. He called that loop feedback. Not in the casual sense people use it today. In the precise mathematical sense. A signal goes out. The result comes back. The system compares the result to the goal. The gap between them drives the next action. The loop closes. That mechanism, exactly as Wiener described it in 1948, is what runs inside every thermostat, every autopilot, every cruise control system, and every AI training loop on the planet right now. When GPT-4 learned to answer questions better, it was doing feedback. When AlphaGo learned to play Go, it was doing feedback. When a self-driving car adjusts its steering because it drifted two inches toward the curb, it is doing feedback. The word they all use, the concept underneath the word, the mathematics formalizing the concept, all of it came from one book written by a child prodigy in 1948 who was trying to figure out how to shoot down a plane. The deeper insight was what he proved about living systems and machines. Before Wiener, biology and engineering were treated as completely separate domains. Organisms adapted. Machines calculated. The idea that you could describe both using the same mathematical framework was not just unusual. It was considered a category error. Wiener proved it anyway. He showed that a brain correcting a reaching movement and a missile correcting its trajectory were running mathematically identical control loops. The hardware was different. The math was the same. Living systems and engineered systems obeyed the same laws once you understood what those laws actually were. He named the field after the Greek word for steersman. Kubernetes. Cybernetics. The person who holds the rudder, reads the water, and adjusts constantly to hold a course through a current that is always pushing the ship somewhere else. That is the mental image he wanted. Not a machine that executes instructions. A system that responds to its own results. The third thing he did is the part almost nobody connects to modern AI. In 1948, Wiener spent an entire chapter of Cybernetics warning about what would happen when machines that learn from feedback were given control over consequential decisions. He described the displacement of workers not as a distant possibility but as a near-term certainty. He wrote about the ethical risks of building systems that optimize for measurable proxies of human values rather than actual human values. He described in plain language what alignment researchers today call Goodhart's Law without using that name, 25 years before Charles Goodhart published anything. He was a mathematician in 1948 writing about problems that AI safety researchers are still trying to solve in 2026. The book is dense in places. The equations are real and the sections on statistical mechanics require actual attention. But Wiener knew this, which is why in 1950 he published The Human Use of Human Beings, which is the same book with all the math removed. Same ideas. Same warnings. Written for anyone who reads English. That second book has been in print for 75 years and almost nobody in tech has read it. Wiener died in 1964 at a conference in Stockholm. He collapsed mid-conversation between sessions. He was 69. He did not live to see a personal computer. He did not live to see the internet. He never saw reinforcement learning, neural networks, or the AI systems that run almost entirely on the mathematical architecture he designed while trying to solve a World War II gunnery problem. Every AI lab in the world today is building systems that run on his framework. Almost none of the people building those systems know his name. The field he founded, cybernetics, mostly disappeared as a word. The ideas did not disappear. They dissolved into every other field. Control theory. Cognitive science. Computer science. Neuroscience. AI. They each took a piece of what he built and called it their own terminology. The word that survived is the one that proves he invented it. Feedback. You use it every day. You use it in code reviews, in meetings, in conversations about AI performance. Every time you use it in the technical sense, meaning a signal that closes a loop between output and goal, you are using the exact definition Wiener wrote down in 1948. He gave the word its meaning. Most people using it have never heard of him. The Human Use of Human Beings is free on archive. Cybernetics is in print and available anywhere books are sold. His major essays are in academic archives at no cost. The man who built the foundation of modern AI was writing about its dangers before the first commercial computer existed. Most people building AI today have never read a word he wrote.

I am the Managing Director of Workforce Transition at a consulting firm that bills $14,200 per day and I am currently advising two clients, in two different industries, running the same playbook from the same deck I built in January, and neither knows about the other. Client A is GitLab. Client B is General Motors. GitLab makes software for people who make software. General Motors makes cars for people who can't afford cars. Both companies, in the same week of May 2026, announced they are replacing their human employees with artificial intelligence products that did not exist when those employees were hired. I built the deck. The deck has 44 slides. Slide 1 is titled "The Agentic Opportunity." Slide 44 is titled "Implementation Timeline." Slides 2 through 43 are the reason I own a house in Darien. GitLab did it with vocabulary. Their CEO published a blog post called "Act 2" on May 7 announcing that the company's six values (Collaboration, Results for Customers, Efficiency, Diversity Inclusion & Belonging, Iteration, Transparency) were being retired and replaced with three: Speed with Quality, Ownership Mindset, Customer Outcomes. I helped write the new ones. Not directly. My firm was not retained for the values work. But I sold the Chief Culture Officer the framework three months ago at a dinner in the Marina where she described the old values as "aspirational scaffolding" and I said, very carefully, that aspirational scaffolding is a liability once the building is up. The building, in this metaphor, is a $1 billion ARR company whose stock has declined 82% from its peak. The scaffolding, in this metaphor, is the 2,000-page public handbook that attracted the employees who are now being told they have eleven days to volunteer for termination or wait until June 1 to learn whether they've been involuntarily selected. The rubric for who stays and who goes contains six dimensions. I know this because I reviewed a draft in March when my associate flew to San Francisco for a "culture alignment session" that was billed as strategic advisory. Two of the six dimensions are "AI fluency" and "agentic mindset." These terms did not appear in any GitLab job description before January 2026. They now determine employment. An engineer who maintained GitLab's CI/CD pipeline for four years without incident — four years of uptime, four years of deployments, four years of the infrastructure that generated the $955 million in revenue the CEO celebrated on the earnings call — may score lower on "agentic mindset" than a new hire who completed a twelve-week certificate in prompt engineering from a program that itself has existed for fewer weeks than the engineer has years of tenure. General Motors did it with spreadsheets. Monday morning, May 11. Badge deactivation at 5:47 AM Eastern, building access at 5:48, VPN credentials at 5:49. Six hundred IT workers across twelve states. The distribution across twelve states was not arbitrary. Each state has a WARN Act notification threshold. Six hundred distributed across twelve states falls below every threshold. The workforce analytics team that designed the distribution model was not among the six hundred terminated. The skill of distributing layoffs across jurisdictions to avoid legal notification requirements is, apparently, an AI-native competency. GM posted 83 new positions the same week. The job descriptions require "AI-native development, data engineering and analytics, cloud-based engineering, agent and model development, and prompt engineering." I reviewed them at my client's request. Several describe roles that the terminated employees were already performing under different names. One posting, Senior Data Integration Architect, is identical to a role held by a woman in their Austin office who was terminated at 5:47 AM Central. She held the position for nine years. The new posting requires three years of experience with large language models. Large language models have existed in commercial deployment for approximately three years. The requirement is mathematically designed to exclude anyone who learned their skills before the technology existed. Which is everyone they just fired. Here is where the deck earns its fee. Slide 17 is titled "The Vocabulary Bridge." It is the most important slide in the presentation. It shows how to construct a lexicon of new competency terms ("AI fluency," "agentic mindset," "AI-native development") that describe existing work in language the existing workforce cannot claim. The vocabulary does not change the job. It changes who is qualified for the job. A senior IT administrator who managed SAP infrastructure processing $185 billion in annual GM revenue for fifteen years is not "AI-native." A twenty-six-year-old with a GitHub portfolio of LangChain wrappers is. The fifteen-year veteran did the work. The twenty-six-year-old has the words. My deck converts one into the other. That is the bridge. GitLab Duo, their AI agent platform, reached general availability on January 15, 2026. Seventeen weeks ago. They are restructuring their entire company around a product that has existed for seventeen weeks. GitHub Copilot has 20 million users and 4.7 million paid subscribers across 90% of the Fortune 100. Cursor reached $2 billion in annualized revenue in February. GitLab's competitor advantage in the "agentic era" is that they are willing to fire more people faster in service of a product that has been generally available for fewer days than their voluntary separation window has hours of anxiety. General Motors spent $10 billion on Cruise, their autonomous vehicle division. Cruise's signature achievement was a robotaxi that struck a pedestrian in San Francisco and dragged her twenty feet. The DOJ fined them $500,000. They settled with the victim for approximately $10 million. They killed the division in December 2024. They then wrote down $7.6 billion in EV losses. They then pivoted back to gasoline. They then announced the 600 IT layoffs for insufficient "AI skills." The AI they built cost $10 billion and injured a woman. The AI skills they're hiring for cost a twelve-week certificate. The employees they fired had fifteen years of keeping $185 billion in revenue processing without dragging anyone through an intersection. Meanwhile — and this is the part where I earn the second half of my fee — GM was simultaneously settling a $12.75 million fine with the California Attorney General for selling the precise GPS coordinates, hard braking events, and real-time driving speeds of 8 million OnStar subscribers to Verisk Analytics and LexisNexis, who used the data to raise those drivers' insurance premiums. GM's privacy policy explicitly stated they did not sell driving data. They sold driving data for four consecutive years. The fine was $12.75 million. The revenue was $20 million. The margin on collecting behavioral telemetry from 8 million of your own customers while the glove compartment manual said otherwise was 64%. The terminated employees' median salary was $95,111. Mary Barra's compensation was $29.9 million. The ratio is 310 to 1. The 1 was just reclassified as "not AI-native." I present these two clients to my partners every Thursday in a meeting we call "Transition Pipeline Review." I present them on the same slide. The slide has two columns. Left column: GitLab. Right column: General Motors. The headers are identical. "Legacy Workforce," "Skills Gap Narrative," "Vocabulary Bridge Deployed," "Separation Timeline," "Replacement Requisitions." The numbers differ. The structure is identical. The structure is always identical. I have seventeen clients in the pipeline. Nine are in technology. Four are in manufacturing. Two are in financial services. One is in healthcare. One is in defense. All seventeen are on slide 17. All seventeen are building a vocabulary bridge. All seventeen are replacing employees who have skills with employees who have words. GitLab's CEO wrote: "Software will be built by machines, directed by people." I read that sentence in a meeting where we were reviewing the rubric for determining which people would be directed out of the company. GM's Chief Product Officer arrived from Aurora, the autonomous trucking startup, to "consolidate disparate technology businesses." Three top software executives departed within six months. Their LinkedIn profiles say "exploring new opportunities" in the same font GM's privacy policy used to say "we do not sell your driving data." Bill Staples's compensation at GitLab was $39.1 million in FY2025. His change-of-control payout is modeled at $47.4 million. Mary Barra's was $29.9 million. Combined: $69 million for two executives presiding over a restructuring that will remove an undisclosed number of humans from payroll and replace them with products that are, respectively, seventeen weeks old and responsible for $10 billion in losses plus one woman dragged through a San Francisco intersection. An anonymous GitLab employee posted on Hacker News: "The employees can have some anxiety until then. As a treat." A GM facilities team filed a maintenance request about moisture on the lobby tables on restructuring mornings. The Warren, Michigan campus has a Panera Bread that opens at 5:30 AM on days when badge deactivations begin at 5:47 AM. The Panera does not know why its hours change. My firm does. We have an agreement with their regional manager. The muffins are complimentary. Slide 17 has a footnote. The footnote says: "Vocabulary Bridge deployment should precede workforce action by 60-90 days to establish institutional legitimacy of new competency framework." GitLab introduced "AI fluency" in January. The restructuring was announced in May. Four months. GM posted "AI-native" job descriptions the same week as the terminations. That is too fast. That is not what the deck recommends. GM skipped the legitimacy window. They went straight from vocabulary to separation without the 60-day buffer that allows HR to say, in the separation meeting, "we communicated these expectations in Q1." I flagged this in my Thursday pipeline review. My partner said, and I am quoting: "They'll be fine. Nobody sues over a word." My deck has been purchased by seventeen companies. The aggregate headcount affected across all seventeen is approximately 14,000 employees. The aggregate revenue of my practice from these engagements is $11.2 million. The per-employee cost of my advisory services works out to $800 per person displaced. That is less than the Panera muffin budget at GM's Warren campus annualized across restructuring days. I have a copy of GitLab's original values poster framed in my office. It says CREDIT: Collaboration, Results for Customers, Efficiency, Diversity Inclusion & Belonging, Iteration, Transparency. I purchased it on eBay from someone whose seller name is "gitlab-alum-2024." I keep it the way a surgeon keeps an X-ray of a interesting case. Not for sentiment. For reference. Slide 44 is titled "Implementation Timeline." It contains a Gantt chart. The Gantt chart has seventeen rows, one per client. Each row has four phases: Vocabulary Introduction, Competency Reassessment, Workforce Action, Replacement Hiring. The phases overlap. They always overlap. The vocabulary is introduced while the competency reassessment is being designed. The reassessment is completed while the workforce action is being calendared. The replacement hiring is posted while the terminated employees are sitting in a Panera at 5:48 AM wondering whether "AI-native" was a term that existed when they were hired. It was not. That is the bridge. That is the product. That is slides 2 through 43. The agentic era is not a technological shift. It is a vocabulary shift. The technology is seventeen weeks old or $10 billion underwater or dragging someone through an intersection. The vocabulary is what my clients are buying. The vocabulary is what makes a fifteen-year SAP administrator into a "legacy workforce" and a twelve-week prompt certificate into a "transition hire." The vocabulary is the product. I am the vendor. The deck is $14,200 per day. The agentic era starts on slide 1 and ends on slide 44 and in between is every employee who built the thing now being renamed to exclude them. I bill monthly. Net 30. The invoices are paid on time. The employees are not.

# Decision Points in AI Agent Development # Trace Sampling Rate 🎯 The Hook Are you recording every single agent trace? Or none at all? The right question isn't "all or nothing" -- it's "what deserves full recording?" A single AI agent request generates thousands to tens of thousands of tokens worth of trace data. Record everything and observability costs explode. Record nothing and debugging becomes impossible. Conditional sampling solves this dilemma. 📋 Overview Trace sampling rate controls what proportion of agent execution traces -- each LLM call, tool execution, and decision step -- are collected and stored. At 100%, every request is traced. At 1%, only 1 in 100. AI agent traces are qualitatively different from traditional web service logs. Including full prompts, full outputs, tool arguments and return values, and intermediate reasoning steps, a single request can generate tens to hundreds of KB of data. Moreover, LLM outputs are probabilistic, so the assumption that "replaying the same input reproduces the same behavior" does not hold. This makes traces irreplaceable for post-incident analysis. 🔍 Decision Points Sampling rate balances accountability and cost_sensitivity, but the most critical design choice is conditional sampling 🎯 Rather than a uniform probability, vary the rate based on request attributes. Priority order for sampling decisions: 1. Requests with errors/exceptions → 100% (mandatory) 2. Requests with HITL (human-in-the-loop) events → 100% 3. Requests involving high-risk operations (side effects, irreversible) → 100% 4. Requests exceeding P95 latency → 100% 5. Requests exceeding cost thresholds → 100% 6. Successful requests → sample at 1-10% 💡 Key Details Reference values 📊 - Error/exception occurred: 100%. Essential for debugging non-reproducible failures - HITL triggered (human approval/escalation): 100%. Required for post-hoc verification of approval decisions - High-risk operations (money transfers, data deletion): 100%. Mandatory for auditing irreversible actions - Latency exceeding P95: 100%. Needed for root cause analysis of performance degradation - Successful and low-risk: 1-10%. Sufficient for statistical quality monitoring - Dev/staging environments: 100%. Full recording where cost isn't a concern Control trace granularity independently from sampling rate 📦 Even for fully recorded requests, store full prompts in the cold tier and metadata (model name, token count, latency, status) in the hot tier. ⚖️ Trade-offs Too low a sampling rate makes failure reproduction impossible 🔍 LLM outputs are probabilistic -- replaying the same prompt won't necessarily reproduce the same error. You also risk missing gradual quality degradation, failing audit requirements, and delayed detection of cost anomalies. Too high a rate and observability costs can rival or exceed production LLM call costs 💸 Performance impact from synchronous trace collection, PII proliferation risk, and signal drowning in noise are additional concerns. Start with a high sampling rate (50-100%) at launch, confirm system stability, then gradually reduce the rate for successful requests. 🛠️ Use Cases Consider tail-based sampling instead of head-based 🔄 Head-based sampling (decided at request start) is simpler to implement, but you can't know upfront whether an error will occur. Tail-based sampling (decided after completion) lets you make decisions based on outcomes, though it requires temporarily buffering intermediate data. Ensure correlation ID (trace ID) propagation is airtight 🔗 In multi-step agent executions, without a consistent trace ID from the first request to the last tool call, you end up with fragmented traces. IDs are especially prone to breaking across async processing and message queues. Combine with hot/cold tier separation for efficiency. Store full trace data in the cold tier for sampled requests and metadata only in the hot tier for everything else -- a practical architecture that balances observability with cost control. #AIAgents# #SoftwareArchitecture#

Gov. Gavin Newsom says he will commute every current death sentence in California to life in prison without the possibility of parole: "I take this action because I firmly believe that state-sanctioned killing is incompatible with California's values."