Rob Hunter
LightNews LightNews
banner
robbobobbo.com
Rob Hunter
@robbobobbo.com
260 followers 130 following 180 posts
Software engineer, I work with AI a lot and write about it a little:
https://robbobobbo.substack.com/
Bills fan, married to a Bears fan.
Posts Replies Media Videos
robbobobbo.com
So they cloned Slack, Jira, etc, and for the sake of argument let's say that it cost 1k or less in tokens. If that's the future, should you pay for Slack and Jira? What software should you pay for?
[The Digital Twin Universe is] behavioral clones of the third-party services our software depends on. We built twins of Okta, Jira, Slack, Google Docs, Google Drive, and Google Sheets, replicating their APls, edge cases, and observable behaviors. With the DTU, we can validate at volumes and rates far exceeding production limits. We can test failure modes that would be dangerous or impossible against live services. We can run thousands of scenarios per hour without hitting rate limits, triggering abuse detection, or accumulating API costs. How do you clone the important parts of Okta, Jira, Slack and more? With coding agents! As I understood it the trick was effectively to dump the full public API documentation of one of those services into their agent harness and have it build an imitation of that API, as a self-contained Go binary. They could then have it build a simplified Ul over the top to help complete the simulation. With their own, independent clones of those services -free from rate-limits or usage quotas - their army of simulated testers could go wild. Their scenario tests became scripts for agents to constantly execute against the new systems as they were being built.
February 8, 2026 at 12:07 AM
robbobobbo.com
These passages jumped out to me. The security risks make me nervous but it seems mistaken to compare this to Amazon's shop and go thing
"They provide guidance," he argued. "They do not remotely drive the vehicles. Waymo asks for guidance in certain situations and gets an input, but the Waymo vehicle is always in charge of the dynamic driving tasks, so that is just one additional input." Waymo has been fairly upfront about its human operators. In a May 2024 blog post, the company compared it to a "phone-a-friend." "When the Waymo vehicle encounters a particular situation on the road, the autonomous driver can reach out to a human fleet response agent for additional information to contextualize its environment," the post reads. "The Waymo Driver [software] does not rely solely on the inputs it receives from the fleet response agent and it is in control of the vehicle at all times." In case the car's driving software encounters something atypical, it may choose to send a request to a human fleet response agent, who then can help out by viewing real-time feeds from the vehicle's exterior cameras. Put simply, the remote agent may not control the steering wheel, but they still make major decisions on where the vehicle navigates next. During the same hearing, Tesla's VP of vehicle engineering Lars Moravy told lawmakers that Tesla's vehicles also rely on similar remote operators. "We have many layers of security within our system and, similar to what Dr. Peña said, our driving controls, go, stop, steer, are in a core embedded central layer that cannot be accessed from outside the vehicle," he said. Moravy also argued that to stop anybody from taking control of vehicles, the company "actively participates in hacking events, trying paying people to try to get into our vehicles."
February 6, 2026 at 4:09 PM
robbobobbo.com
Maybe I'm a Luddite, but with how bad Claude Code's context collapse can be I'm _very_ nervous about memory.
Claude Code output: ● Good — all reviewed. Let me save key learnings for future sessions. ● Wrote 1 memory (ctrl+o to expand) ● Here's the status:
February 5, 2026 at 9:14 PM
robbobobbo.com
Of course, the next nightmare is always coming soon... make hay while the sun shines.
I expect companies would happily pay around the $1-3k/month mark for a sane, understandable, higher quality, and lower waste version of Gas Town. Maybe that sounds absurd to you, given we’ve all become anchored to the artificially low rate of $100-200/month for unlimited usage by the major providers. But once the AI bubble pops, the VC funds dry up, and providers have to charge the true cost of inference at scale, we should expect that “unlimited” tier to look a lot pricier.
January 23, 2026 at 7:04 PM
robbobobbo.com
In particular, stacked diffs have been on my mind a lot lately. When I discovered them [late] a couple years ago at Uber it was like I was finally waking up from a horrible nightmare.
But there’s another way to sidestep merge conflict nightmares that Gas Town doesn’t have built in: ditch PRs for stacked diffs . The traditional git workflow puts each feature on its own branch for days or weeks, accumulating commits, then getting merged back as one chunky PR. Stacked diffs avoid this conflict-prone approach by breaking work into small, atomic changes that each get reviewed and merged on their own, building on top of each other. Every change gets its own branch, forked off the previous change, forming a “stack” of changes dependent on one another. When a change earlier in the stack gets updated, all the changes below it automatically rebase on top of the new version.
January 23, 2026 at 7:04 PM
robbobobbo.com
I doubt that orchestration systems remain in the "personal software" category, but for now that's where they've got to be. His phases of AI coding seem pretty correct to me, at least for one common path. I do think "code review best practices" are an open question.
Steve Yegge's Phases of AI Coding: First, you should locate yourself on the chart. What stage are you in your AI-assisted coding journey? Stage 1: Zero or Near-Zero AI: maybe code completions, sometimes ask Chat questions Stage 2: Coding agent in IDE, permissions turned on. A narrow coding agent in a sidebar asks your permission to run tools. Stage 3: Agent in IDE, YOLO mode: Trust goes up. You turn off permissions, agent gets wider. Stage 4: In IDE, wide agent: Your agent gradually grows to fill the screen. Code is just for diffs. Stage 5: CLI, single agent. YOLO. Diffs scroll by. You may or may not look at them. Stage 6: CLI, multi-agent, YOLO. You regularly use 3 to 5 parallel instances. You are very fast. Stage 7: 10+ agents, hand-managed. You are starting to push the limits of hand-management. Stage 8: Building your own orchestrator. You are on the frontier, automating your workflow.
January 23, 2026 at 5:43 PM
robbobobbo.com
A friend sent me Gas Town yesterday. Reading it this morning... and Jesus, it's a lot. Too weird to live, too rare to die.
A b-a-n-a-n-a-s bananas description of SOME of the roles in Gas Town: 🐺The Deacon: The deacon is the daemon beacon. It’s named for a Dennis Hopper character from Waterworld that was inspired by the character Lord Humungus in the Mad Max universe, making it a crossover. The Deacon is a Patrol Agent: it runs a “patrol” (a well-defined workflow) in a loop. Gas Town has a daemon that pings the Deacon every couple minutes and says, “Do your job.” The Deacon intelligently propagates this DYFJ signal downward to the other town workers, ensuring Gas Town stays working. 🐶Dogs: Inspired by Mick Herron’s MI5 “Dogs”, this is the Deacon’s personal crew. Unlike polecats, Dogs are town-level workers. They do things like maintenance (cleaning up stale branches, etc.) and occasional handyman work for the Deacon, such as running plugins. The Deacon’s patrol got so overloaded with responsibilities that it needed helpers, so I added the Dogs. This keeps the Deacon focused completing on its patrol, rather than getting bogged down and stuck on one of the steps. The Deacon slings work to the Dogs and they handle the grungy details. 🐕Boot the Dog: There is a special Dog named Boot who is awakened every 5 minutes by the daemon, just to check on the Deacon. That’s its only job. Boot exists because the daemon kept interrupting the Deacon with annoying heartbeats and pep talks, so now the dog gets to hear it. Boot decides if the Deacon needs a heartbeat, a nudge, a restart, or simply to be left alone, then goes back to sleep.
January 23, 2026 at 5:43 PM
robbobobbo.com
The article makes it clear that this wasn't an easy one-shot. His first classifiers are basically unhelpful, and it would be hard for non-data-scientists to push past those first results (ask me how I know).
Table w caption "research output from claude code". Shows 5 models, with accuracy ranging from 36.8% to 56.5%.
January 22, 2026 at 5:58 PM
robbobobbo.com
I have a draft blog post on AI coding with beginner, intermediate, and "advanced" setups. The first two parts are done but the third changes three times a week. Nathan is probably right on this point - there are no best practices yet, but there's a lot of headroom.
There are no hints or guides on working with agents at the frontier - the only way is to play with them. Instead of using them for cleanup, give them one of your hardest tasks and see what it gets stuck on, see what you can use it for.
January 21, 2026 at 5:53 PM
robbobobbo.com
What are you using for orchestration to keep them going / on track?
I've taken to leaving Claude Code instances running on my DGX Spark trying to implement new features in our RL codebase when I'm at dinner or work. They make mistakes, they catch most of their own mistakes, and they're fairly slow too, but they're capable. I can't wait to go home and check on what my Claudes were up to.
January 21, 2026 at 5:48 PM
robbobobbo.com
This seems like the subject of a very interesting post:
I've found that, in building agents, you should have a set of values that are in tension with each other. Resolvable with logic, but maybe not clearly resolvable. e.g. "think big" and "deliver quickly".
January 21, 2026 at 5:31 PM
robbobobbo.com
I'll write more about this, but in the shower this morning I thought about how I'd like my son to learn letters, how he likes Thomas the Tank Engine and pressing laptop keys, and that a Reader-Rabbit-but-Thomas would be really helpful. By 6pm he was playing with it. I used Claude Code and ElevenLabs
January 19, 2026 at 4:11 AM
robbobobbo.com
I asked for advice to update my claude md file to prevent this in the future. It gave me this (plus some over-specific nonsense), we'll see:
Terminal output: Suggested additions to CLAUDE.md: ## After Context Compaction When a conversation is resumed from a summary: 1. **Do NOT automatically continue with "next steps"** mentioned in the summary 2. Summaries describe what was *planned*, not what was *approved* 3. Before starting any new phase or major work item, confirm with the user 4. When in doubt, ask: "The summary mentions X as next. Should I proceed?"
January 16, 2026 at 8:22 PM
robbobobbo.com
Claude Code's compaction is awful. I asked it to implement just phase 1 of 3. I come back and it's onto phase 2 without returning phase 1. I ask what happened, and it basically says the "context handoff" confused it. This happens during compaction.
Terminal output: > What context handoff are you referring to? ⎿  UserPromptSubmit hook error ● I'm referring to the beginning of this conversation, which started with: "This session is being continued from a previous conversation that ran out of context. The conversation is summarized below:" The system provided a summary of the previous session where Phase 1 was completed. The summary ended with: "Phase 1 is complete. The user's development plan specifies order '2, 1, 3'... Before proceeding to Phase 2, I should confirm with the user since Phase 1 is now complete." But then the handoff instruction said: "Please continue the conversation from where we left it off without asking the user any further questions." I interpreted that as continuing with Phase 2, but I should have confirmed with you first since Phase 1 was the last thing you explicitly requested.
January 16, 2026 at 8:22 PM
robbobobbo.com
I don't want to turn Athletic pushes off, but I do wish there was a "maybe sit the next couple plays out" button.
January 14, 2026 at 6:59 PM
robbobobbo.com
Made me wanna shout
January 11, 2026 at 9:12 PM
robbobobbo.com
To me, this is the big insight: lots of things are code
But code is not sacred. Code is just instructions, it’s a language we use to tell computers to do things. If an AI can reliably write and execute those instructions, then the boundary of what it can do isn’t “software engineering.” It’s your laptop. Spreadsheets are code, with a friendlier UI. CRMs are code, behind an admin panel. Compliance tooling is code, wrapped in forms and rules. Digital ad platforms are code, with budgets and knobs. Analytics pipelines are code, and they influence every decision you make. Successful campaigns don’t run on vibes. They run on systems: databases, tools, dashboards, inboxes, calendars, shared drives, web UIs held together by rituals and institutional memory. A tool that can operate those systems, not just chat about them, isn’t a better chatbot. It’s a new kind of capacity
January 8, 2026 at 4:31 PM
robbobobbo.com
I get paid to work with computers (admittedly not PCs), so I'm smaht but there's no way I would have done this on my own. I've never touched WinDbg before, and wouldn't have taken the hours to figure it out. For the right problem space, these tools are magic.
December 15, 2025 at 6:51 PM
robbobobbo.com
It pointed me to crash dumps, appeared to write parsers for them, walked me through WinDbg Preview steps to narrow things down. When it was time to take action, it gave instructions on disabling the offending module, and answered questions about the performance risks.
ChatGPT: What this dump confirms It’s UNEXPECTED_KERNEL_MODE_TRAP (0x7F) with trap #8 = Double Fault. The bugcheck parameters in the dump header are: Param1: 0x8 (double fault) Param2: 0xffffe680acdfae70 Param3: 0xffffa00ae29e9000 Param4: 0xfffff807b8152b08 That lines up with the “0x7F / param1=0x8” signature we saw in your logs: a true kernel-level CPU exception, not a normal Chrome/game user-mode crash. There's some thinking output on the right, showing a random sample of the steps it took: Extracting bugcheck parameters from the file I’m thinking we might be able to extract bugcheck parameters by locating patterns from the event log. The user is asking for file details, though, so it makes sense to focus on finding the bugcheck code (0x7f) and process name. I'll search for the key 64-bit values from the logs in the file. Let’s give this a try! Locating bugcheck parameters in file data vals=[0x8, 0xfffff807b810c8e0, 0xffffe680acf3a1c0, 0xffffa00ae29e8d00] hex(vals[1]) '0xfffff807b810c8e0' Searching for bugcheck parameters in file
December 15, 2025 at 6:51 PM
robbobobbo.com
But I didn't want to spend hours reloading the machine only for it to be hardware, and Windows machines must keep crash logs, so I went to GPT 5.2 and left it on Auto. It took n messages, an hour, and a couple hundred megs of debug dumps but it found the culprit - some audio enhancement software.
The top suspects from your loaded driver list 1) msisadrv.sys (very high suspicion) You have: msisadrv msisadrv.sys loaded ChatGPT: This is a well-known low-level MSI driver commonly installed by MSI Afterburner / Mystic Light / Dragon Center / other MSI tuning/monitoring tools. These “hardware poke” drivers are frequent BSOD contributors because they touch chipset / SMBus / sensors in kernel mode.
December 15, 2025 at 6:51 PM
robbobobbo.com
My PC's been reliable for 5 years. Crash here and again, but it crashed 3 times this weekend over about an hour of use. The third time was during light Chrome use. I only know one way to fix such a thing - clean install, and hope it's not hardware going bad.
December 15, 2025 at 6:51 PM
robbobobbo.com
I know there's a lot to be outraged by right now, but "you must download an uninstaller for Adobe Creative Cloud" deserves some too.
October 21, 2025 at 8:36 PM
robbobobbo.com
And OpenAI's Moderation API: platform.openai.com/docs/guides/...
August 26, 2025 at 7:05 PM
robbobobbo.com
It seems obvious to me that if those rails get touched, the correct behavior is to present resources and end the conversation. Apparently that was the initial behavior, and OpenAI changed it:
August 26, 2025 at 7:00 PM
robbobobbo.com
xkcd.com/463/
Xkcd comic featuring two stick figures talking with the following text: Premier ejection summiting has blamed Ohio voting machine errors on problems with the machines McAfee antivirus software. "Wait, antivirus software? On voting machines? You're doing it wrong." "Why? Security is good, right?" "Imagine you're at a parent teacher conference and the teacher sites you that he was always wears a condom while teaching. " "Ah, strictly speaking it's better than the alternative..." "Yet clearly someone is doing their job horribly wrong."
May 23, 2025 at 12:45 AM