Estimated reading time at 200 wpm: 6 minutes
Date: 8 February 2026. System: Scan 3XS Laptop (Clevo X56xWNx Architecture), Processor: Intel Core Ultra 9 285HX (Arrow Lake), RAM: 64GB DDR5 RAM, DRIVES: 3TB NVME, Graphics: NVIDIA GeForce RTX 5070 Ti Laptop GPU (Blackwell). In simple terms, this just means no-nonsense very high end. Luddites will be deterred by all the ‘jargon’ in here. But actually this post is not really about technology. It is about where AI is strong and where it is weak.
Whether or not you agree our Fat Disclaimer applies
1. The Core Problem: “Lights Out”
The crisis began with the “Lights Out” bug—a catastrophic failure where the laptop would enter an S0 sleep state [you don’t need to know what that is!]. On reawakening any touch of the keyboard would cause a crash; no blue screen – just sudden shutdown – lights out. By examining error logs with AI assistance, this was traced to NVIDIA Driver 591.86 (January 2026). Loads of people are in tears over the last several months with similar problems caused by Nvidia’s incompetence across several drivers – for people with my sort of high end computer. For all it’s AI might, Nvidia hasn’t found a solution. They’re probably too busy partying – drunk on their trillions from stock-market victories.
- The Conflict: The 591.86 driver introduced a regression in PCIe power-state transitions. On Arrow Lake systems, this caused the motherboard’s Embedded Controller (EC) to “latch” a power fault.
- The Result: To protect the hardware from potential voltage spikes during a failed wake, the motherboard gated power to the GPU. This led to the “No GPU found” error, making the RTX 5070 Ti invisible to Windows Device Manager and installer packages.
2. Driver Version Failures
Before stabilisation, multiple driver versions were attempted, each failing for specific reasons:
- 591.86 (The Culprit): Forced via Windows Update; caused the initial power-state crashes and hardware disconnects. But this issue has been going on since August 2025 through several driver updates. I assumed in Jan 2026 they would have sorted it. I was wrong.
- AI suggested a 32.0.15.7284 (Dell-hosted): Failed with Error 1603. This was a branding mismatch. The installer recognised the motherboard was not Dell-proprietary and aborted the payload before it could reach the GPU. Pretty dumb idea. Then AI pursued this by trying to decompile the Dell driver. I spotted this rabbit hole.
- NVIDIA Advanced Search (Redirect Loops): NVIDIA’s own site failed to serve older “Production Branch” drivers (like 573.24) for the 50-series, instead redirecting to newer, unstable builds or irrelevant support pages.
Arrow Lake computers utilise modern ACPI power management standards, primarily relying on S0ix (Modern Standby) for rapid wake times, while supporting traditional S3 (Suspend to RAM) and S4 (Hibernation) states. They feature advanced per-core and package-level C-states to minimize power consumption when idle. Blackwell GPUs (RTX 50-series) include faster, more efficient deep sleep state transitions.
Key Sleep and Power States for Arrow Lake
- S0ix (Modern Standby): The primary low-power state for mobile Arrow Lake systems, offering quick resumes similar to a smartphone.
- S3 (Suspend to RAM): Traditional standby, saving power while keeping system context in memory.
- S4 (Hibernate): Saves all data to the drive and powers off, ideal for long-term suspension.
- C-States (Core/Package Idle): Deep sleep states (𝐶3, 𝐶4, etc.) are used to gate power to idle CPU cores, reducing energy consumption.
- P-States (Performance States): Used to adjust voltage and frequency while active.
What that means is that if S0 is in trouble due to driver issues, there is trouble down the line.
3. The Turning Point: BIOS & Hardware Re-Discovery
When the hardware disappeared entirely, standard software fixes were useless. The following hardware-level steps were required to bring the GPU back into “sight”:
- EC Reset (The 60-Second Hold): Drained flea power from the motherboard to clear the “latched” power fault. This part was based on sound knowledge by the AI.
- BIOS Display Mode Adjustment: The system had defaulted to Integrated GPU mode as a safety precaution. By manually toggling this to Discrete GPU, the motherboard was forced to re-energise the PCIe lane for the NVIDIA card.
- Result: The GPU reappeared in Device Manager as a Microsoft Basic Display Adapter (VEN_10DE), confirming the silicon was alive but lacked a driver “identity”.
4. The AI “Rabbit Hole” Analysis
Progress was significantly delayed by the AI’s tunnel vision and reliance on finding specific “Golden” driver versions (like 573.24) that were no longer indexed or accessible via public URLs.
- The Error: The AI prioritised theoretical “perfect” versions over immediate practical solutions. By providing dead links and suggesting OEM-specific packages (Dell) that were guarded by branding checks, the AI wasted critical time on search-loop redirections and incompatible software wrappers.
- The Lesson: On cutting-edge hardware (Blackwell/Arrow Lake), AI guidance should pivot immediately to stable builds already in the user’s possession rather than hallucinating paths to delisted historical versions.
4.1. Human Identification: The ‘Best Guess’ Exit
The stalemate was broken only when the human identified the AI’s repetitive failure loop and pivoted to a “Best Guess” strategy using locally available assets.
- Cutting the Loop: While the AI was stuck in a search-and-redirect cycle, the human correctly recognised that the path forward lay in the local downloads folder (577.00, 576.52, 572.83).
- The Logic: The human applied a practical heuristic—577.00 was a production-branch driver released after the initial Blackwell bugs were patched but before the catastrophic 2026 update.
- The Result: This human-led override bypassed the AI’s “rabbit holes”, allowing for a successful installation and hardware re-identification within minutes, proving that local, known-good files are superior to theoretical web-searched “gold” versions.
5. Final Resolution: The 577.00 Install
The recovery was finalised when the user overrode the AI’s search loops and made the executive decision to use 577.00 as the “Best Guess” for stability.
- Installation Method: Performed via the NVIDIA app with a rejection of the “Studio” update to prevent further version-bloat.
- Verification: The OEM Control Center now successfully reports GPU clock speeds (180 MHz) and temperatures (43°C), proving the driver-to-hardware power handshake is restored.
- Lockdown: To prevent a recurrence, Group Policy was enabled (
gpedit.msc) to explicitly exclude driver updates from Windows Update.
6. Observation Period
The system is currently running in Discrete GPU mode. This is the most stable state as it eliminates the Intel/NVIDIA switching logic. We are now monitoring for:
- Reliable wake-from-sleep behaviour.
- Absence of “nvlddmkm” timeout errors in Event Viewer.
- Thermal stability under sustained workstation loads.
7. Conclusion
AI is a sound brute force assistant. I could not have worked out how to adjust the BIOS on my own.
But thereafter the familiar pattern of going down rabbit holes emerged again. It saw each roadblock as something to get around. AI avoided making the best guess, when given options. I had to push. I took control and made the best guess. Then there was progress with the 577 driver, as imperfect as it is – but less imperfect than 591 latest drivers (for my purposes). I’m no gamer. I just needed a reasonably well working graphics card and a computer that was not crashing after coming out of hypersleep. Overnight and well into Feb 9th 2026, all is well. No more crashes. Now I wait to see if the same issues emerge again.
