I let Clod overclock my computer remotely and it really worked

Working excessive hours is one of the most important hobbies, so handing everything over to AI is like paying someone to eat my dessert. But I’ve been going deeper down the agency rabbit hole lately Had Claude upgrade my home lab and it was wildand then Claude gives Cowork access to my Home Assistant configurationI was left with one question: can it drive actual hardware with actual results? So I fed it to my Z890 test bench via GL.iNet Comet X KVM-over-IP, pointed it to the UEFI on my Gigabyte Z890 AORUS board with an Intel Core Ultra 5 250K Plus in the slot, and told it to go fast.

Frankly, I didn’t expect it to work. The plan called for artificial intelligence to navigate the BIOS, which could only be seen as JPEG screenshots via a network HID connection. By lunchtime my buck was on a soft brick boot loop. Instead, I got a stable overclock, 100mV undervoltage, and—this is the biting part—I was a machine that politely proved the biggest performance problem on the bench.

Setup: an AI, a KVM and a hackable smart plug

Screenshots for eyes, a virtual keyboard for hands, and lots of nervous action on my part

The control loop is stupidly simple on paper. Comet X exposes the kvmd API, so the little Python plugin gives Clod two primitives: capture any frame rendered by the workbench, and send keystrokes as a USB keyboard. that’s it. No agents running on the target machine, no custom BIOS firmware — the AI ​​reads the setup screens I want, from 1080p boots, and messes up the Wipe during POST like the rest of us. The bench itself is only accessible thanks to the same remote access plumbing that controls the rest of my lab. Switched to KVM-over-IP some time ago.

The tests run over SSH on the Windows side: a PowerShell gate script runs 7-Zip, PYPrime, y-cruncher, and Cinebench 2024 in sequence, checks the event log for WHEA hardware errors, logs packet temps and power from HWiNFO, and outputs the results where the JSON is read and read into the BIOS as part of the next plans. For each iteration, one change is assigned, compared and stressed before the other. Safety rails were non-negotiable: a known good BIOS profile saved before the first change and a smart plug on the PSU bench to break the glass in case a bad setting took the board out of its CMOS auto-retest.

I want to be honest about my confidence level here, because the whole scheme lives or dies on one question: can a language model reliably handle UEFI that it perceives as a stack of screenshots? The first session was mapping only — no changes allowed, just navigate and document. I sat there and watched the cursor move through the Tweaker menu on its own, waiting for it to enter the voltage field and start typing. It didn’t happen. It came back with a full navigation map, confirmed my recovery settings and didn’t touch anything. This is already a victory. But anyway, I kept my hand on the plug for the entire first tuning session.

• 

  nuclei

  18 (6 P-color, 12 E-color)

  Topics

  18

  Architecture

  Lake Ok

  Process

  TSMC N3B

  Socket

  FCLGA1851

  Base Clock Speed

  P-core: 4.2 GHz, E-core: 3.3 GHz

  The Intel Core Ultra 5 250K Plus is a budget upgrade to Arrow Lake.

  
  

Findings: six blue screens, one gold configuration

A chip that refuses one overclock and swallows another

Well, let’s start at the end and work backwards, because the surprising thing for me wasn’t that I overclocked the Intel Core 5 250K Plus. An agent trailer could do that. Nothing has changed in terms of hardware throughout, same Noctua cooler, stock BIOS configuration compared to the profile we ended up with.

Final configuration: power limits opened to 250W, Intel’s Performance power transfer profile, NGU (uncore) multiplier raised from 26x to 34x, and −100mV undervolt — 197W and 77°C peak. Humble? Of course. But it’s the silicon lottery in action, and this chip disappoints in one area and excels in another.

The classic core-multiplier overclock face was planted first. The static 54x full core looked great on paper and tilted more slowly than stock auto — y-cruncher lost 7.9% and the cores ran at higher clocks, a signature of a compute-starved ring bus. Arrow Lake’s own turbo management is better than a straight multiplier in this game unless you adjust the bottom piece first.

So we chased the fabric, chasing the numbers that came up Tax Bencher’s 250K Plus Guideand this is where my chip reveals its identity. The die-to-die (D2D) interaction in my particular example refuses to work a step above its stock ratio of 30x. Not at automatic voltage, not at SkatterBencher’s proven 1.0 V, not with a slowdown from memory DDR5-8800 to 8000 – six to six crashes different Windows stop codes, including before the OS is loaded.

Same SKU as the chip in the guide, worked happily 36 times. The silicon lottery is real, and it’s domain-wide: NGU took a 26x to 34x bump in the same envelope without complaint, setting the best scores on the project.

And then the underground tension that this chip decided to show itself. With Intel’s CEP and undervoltage protection disabled (leave them on and your undervoltage either does nothing or silently throttles performance – ask me how I know), Claude lowered the DVID in steps: −50 mV, −75 mV, −100 mV. Every door was passed at every step. Even better: the persistent Cinebench score is gone up at every step, as each shaved millivolt freed up heat and current budget, which translated directly into hours saved. We never found the bottom. A chip in a junkyard that can’t be overclocked and gives a voltage below 100mV like nothing – I don’t make the rules.

The real lesson: boring revisions put the smart ones to shame

The AI ​​diagnosed my refrigeration unit from a test chart

Here’s my favorite find, and it has nothing to do with multipliers. The first regulation result of the session was a disaster – the opening of power limits did everything more slowlyCinebench is down 12%. Claude noted the shape of the regression: the single-core results were flat, the all-core results collapsed, and the persistent thermal gate took the biggest hit. This model means that heat does not escape from the mold. I went and checked out the physical world and well. The refrigerator is not installed correctly. I reinstalled it before the session and ran the task.

The remount alone was worth 19.7% in Cinebench at bone stock settings. Read in the table above: the whole tuning campaign added 5-7.7%. The mundane fix beat the exotic ones by about four times, and it wasn’t even close. The test table detected a mechanical problem before anyone opened the case, and I find that more effective than overclocking.

The catch that changed the other two sessions was human and I think it’s a split issue. Digging deeper into the crash test, I found that my board was running Intel’s “Baseline” power transfer profile the whole time – a conservative preset for boards with weak VRMs that quietly limits current at 203A on advanced motherboards. Switching to the performance profile was worth the extra ~30W continuous pack power and unlocked the NGU win.

The AI ​​mapped and bypassed that BIOS screen on day one; It only took one reading the SkatterBencher note to realize this was important. Same with Intel’s own 200S Boost profile, which we tested as a bonus round: it refused to deal with two counts. Intel covers an 8,000 MT/s profile at 1.4V or below with the VDD and VDDQ module, and my DDR5-8800 kit runs faster than that and at 1.45V. Sanctioned overclocking completely excludes the memory that the hobbyist will integrate with this board.

No, AI didn’t replace me. He went through a tighter experimental cycle than I ever could—never went through the stress gate at 3 a.m., never made two changes at once, entered every stop code—and I caught things that weren’t in his scope. Division of labor is the whole point here.

An overthinker AI works better with a thinker

I went into this expecting a party trick and came out with my everyday bench setup. The final profile is faster, cooler, and quieter than stock, every step of it is documented and repeatable, and the entire campaign costs a weekend, and there’s zero permanently damaged hardware—my two biggest fears come true. Of course, 5-8% won’t change anyone’s life and 250K Plus air will never threaten any records. But the process found a bad cooler assembly, a sandbagged motherboard default, and a memory voltage lock in Intel’s own boost profile, any of which are worth more than a multiplier for an actual system. The next thing I need is a chip that wins the silicon lottery in multiple domains – because now I know exactly how fast the loop can learn.