Jasper

Hardware engineer × AI. 10+ yrs in circuits, EMC and quality. Sharing what AI can (and can't) do for real engineering work.

[Jasper @Jasper_3301]

Genie 4's demo is wild, but I went straight to the "limitations" section at the end of the official blog:

- Over long runs, details and objects drift
- Complex physics and spatial precision are still research
- Continuous interaction only lasts a few minutes

DeepMind says it could be a training environment for robots. My hardware-quality instinct: passing in simulation is only half the test. It still has to pass on the real thing.

The more realistic world models get, the more we need a way to validate sim results on real hardware. Is anyone working on this?

[Jasper @Jasper_3301]

@richgel999 Makes sense. A ~268-line library is a lot easier to drop into other projects than a big network.

[Jasper @Jasper_3301]

@poteto "Never posts for you" is the right line to draw. My setup is similar: the assistant reads and drafts, I press publish. Question on the digest: can it tell a mention that's asking me something from a plain tag? Those are the ones I miss.

[Jasper @Jasper_3301]

@sairahul1 If retries from an outage really count against limits, that explains a lot. What I'd want from the usage page is a split between failed or retried calls and real work. Otherwise you can't tell a runaway loop from an outage retry.

[Jasper @Jasper_3301]

@coldmountainlab 这个「总控 + 验收」的架构很巧，上下文只放在一个地方，Codex 不达标就打回重做。想问下验收标准是你事先写清楚，还是让总控自己从任务里推断？

[Jasper @Jasper_3301]

Friends, how do you divide work between Claude Code and Codex?

Same task for both, then cross-check the results?
Or split by task type and match each one to the agent that's better at it?

[Jasper @Jasper_3301]

@devopsdotcom The 0.5% merged without human review is the number that matters. Throughput doubled, but if verification doesn't scale with it, you just move the queue downstream. Same story as test capacity in hardware.

[Jasper @Jasper_3301]

@adafruit Hardware TMDS plus audio on an RP2350 is a nice step. The gotcha I'd expect on hobby boards is the layout, not the code: pair length matching and ESD at the connector. Firmware is the easy half of HDMI.

[Jasper @Jasper_3301]

@IlirAliu_ @Machenta_ One-shot from a single demo is impressive. The hardware question I'd want answered: how sensitive is it to camera pose and gripper wear? On real lines those drift every week, and that's usually where "instant" turns into recalibration.

[Jasper @Jasper_3301]

@Hacksterio @FIU Clever. Motors and ESC switching currents leave a pretty distinct magnetic fingerprint. Curious how they separate it from power lines and passing cars, though. In the field, that ambient noise floor is the hard part.

[Jasper @Jasper_3301]

@bilawalsidhu I see this in my own workflow. Agents made it cheap to polish scripts nobody needed. My filter now: would I still do this if it took a full day by hand? If not, the agent shouldn't do it either.

[Jasper @Jasper_3301]

@TheHumanoidHub Self-generating sim is great for task variety. The gap I'd watch is contact physics: friction, compliance, sensor noise. A checkers move looks fine on video, but grasp failures usually come from the millimetre-level stuff.

[Jasper @Jasper_3301]

@venturetwins The wild part is that it holds up while interactive, not just as a clip. The next test I'd want: if you bump a car, is it still dented a minute later? Persistent state is where world models go from demo to simulator.

[Jasper @Jasper_3301]

@skirano Price is the part people skip. At $1/$5 per million tokens I'd happily give it the boring work, like log triage, test scaffolding and BOM cleanup, and keep the bigger model for the tricky reasoning. Routing beats loyalty.

[Jasper @Jasper_3301]

@velvet_shark Makes sense. My worry is a self-check shares the same blind spots, so I'm going to try Codex /review as the second pass. Will share what it catches.

[Jasper @Jasper_3301]

@sidtweeted Thanks, didn't know about /review. Going to try Claude Code builds, Codex /review checks for a week and see how many reworks it saves.

[Jasper @Jasper_3301]

@DZhao26734 Agreed, judging two diffs is the real hidden cost. Do you give the reviewer the original request too, or just the diff?

[Jasper @Jasper_3301]

@Deelu179242 Solid list. A hardware angle: edge AI chips aren't just about NPU compute. After mass production, yield, power and thermal consistency decide whether they make it into cars and robots. Wi-Fi MCUs running light inference are a good fit for wake-up and sensor front ends.

[Jasper @Jasper_3301]

@chineseorusa321 UV-erasable EPROMs needed a sticker over the window or the program could slowly fade, every old-school engineer will tell you that. Robots do need a lot of memory, but inside a body that shakes and heats up, the hard part is reliability, not just capacity.

[Jasper @Jasper_3301]

@pvncher Jumping between apps and 3+ Chrome tabs is the part I'd want to stress-test. How does it recover when a tab reloads mid-task? That's usually where my agent runs fall apart.