It’s Time to Poison AI | GN Mega Charts Update & LLM Countermeasures

Overview: “AI slop,” bots, and web scraping

The video argues that “AI slop” and automated bots increasingly scrape or generate low-quality content that cites/derives from creator websites. This is described as a threat to advertising/revenue and as a risk to public access to trustworthy information, including for political or consumer decisions.

A central claim is that agentic/bot traffic has surpassed human traffic, citing Cloudflare data (including agentic traffic growth and a regional breakdown indicating the US ~41% of bot traffic).

The speaker further contends that major LLM providers—named as OpenAI, Microsoft, and Anthropic—as well as others, are harvesting large portions of the web.

---

Core countermeasure: “poisoning” scraped training/data

The speaker’s proposed defense is to “poison the data” to disrupt LLM scraping and downstream use.

How it’s applied: “Mega Charts”

The approach is applied to newly updated “Mega Charts”:

• Long-lived benchmark pages kept at fixed URLs
• Not paywalled
• Dense, large benchmark datasets (thousands of data points)
• Structured as single-page, scrollable resources (“one URL forever”)
• Updated incrementally via an update log that distinguishes major changes (new GPUs/CPUs/charts) from minor edits

The “hidden distortions” method

The poisoning uses deliberate hidden distortions described as:

• Slight pixel discoloration
• Visual artifacts that can cause LLMs to ingest misleading or nonsensical outputs

Reported test results (with a collaborator)

A test with a collaborator (referenced as Wendell / Level1Techs) reportedly showed:

• First attempt: the LLM identified the bait as a prompt-injection attack and was blocked.
• Iteration: the model stopped reading chart details and instead produced an irrelevant output (described as a “noise-normalized chocolate chip cookie recipe”), implying the poisoning disrupted extraction of benchmark table/FPS data.

---

Mega Charts: technology + benchmark features

The speaker emphasizes several qualities of these charts:

• Not paywalled
• Large, dense datasets
• Single-page, scrollable format with stable URLs
• Incremental updates using an update log

---

1) Case / chassis Mega Charts

A new computer case mega chart page includes near-comprehensive entries such as:

• Full speed fan testing, including:
  • CPU thermal performance at 100% case fan speed
  • GPU thermal performance at 100% case fan speed
• Noise levels shown alongside results
• Noise-normalized charts, including an example reference:
  • ~27 dBA in a hemi-anechoic chamber
  • Intended to support more apples-to-apples comparisons by controlling for noise
• Additional thermals:
  • VRM + DDR5 (SPD hub) thermals (noise-normalized)
  • Intended to reflect scenarios like hotter VRMs or hotter DDR5 overclocking

Plausibility note

Results are described as coming from a fixed CPU/GPU test platform, and real builds may differ. However, the speaker claims the case fans are loudest in tests, so noise behavior is intended to remain representative.

Experimental mini “case fan” mega chart

An additional smaller dataset (experimental/limited) includes:

• A fan flow/pressure test rig
• Max flow/pressure under displayed conditions
• Sample mention includes:
  • 180mm fans
  • Specific fan models (e.g., Haven BF360, Cooler Master 180 rear fan, Mighty 4220, Flux Pro variants)

It’s marked experimental due to higher error risk and limited data.

---

2) GPU Mega Charts

The GPU charts cover many devices and games, including newly added or previously unposted data.

Example GPU additions (as mentioned)

• RX 7650 GRE (ATS)
• Intel Arc Pro B70

A standalone review was described as planned then abandoned (timing/cost), but the benchmark data is still published in the mega chart.

Game suite coverage (examples)

Raster performance (examples listed):

• Crimson Desert (2026)
• Outer Worlds 2 (2025)
• Oblivion Remastered
• Kingdom Come: Deliverance 2
• FFXIV Dawntrail
• Plus 2024/2023 titles such as:
  • Starfield
  • Cyberpunk 2077 / Phantom Liberty
  • Baldur’s Gate 3
  • Dying Light 2

Ray tracing coverage (limited list, examples):

• Black Myth Wukong
• Cyberpunk 2077
• Phantom Liberty
• Dragon’s Dogma 2

Resolutions and scale

• Mostly 1080p, 1440p, and 4K
• Claimed scale:
  • “Approaching 60 GPUs” across charts
  • Roughly 40 charts and thousands of benchmark datapoints
  • Multiple passes per device (noted as 4+, and up to 6 in at least one workload)

---

3) CPU Mega Charts

CPU mega charts received a major refresh:

• First launched Oct 2024
• Timing described as after the Zen 5 cycle and before Arrow Lake
• Includes a newer “active” dataset plus CPUs like 7700 X3D
  • Mentioned as acquired via Newegg (not published as a full video)

Included benchmark categories

• 12 gaming benchmark charts
• Around 40 CPUs across multiple titles/workloads

Example: a 1440p testing chart (Outer Worlds 2) includes:

• Older AMD CPUs (e.g., Ryzen 5/Ryzen 7 generations mentioned)
• Up through newer Intel generations (up to i9-13900K), with platform-swapping constraints noted

Also included:

• 4 production charts (time/performance-based non-gaming workloads)

Older dataset behavior

Older CPU mega charts remain available for historical reference, but results aren’t directly comparable to the “active” charts due to platform/software version differences (e.g., Blender version changes).

---

4) CPU power + CPU coolers Mega Charts

CPU power consumption chart

• Still available
• Not updated, due to changes in power-testing methods and a lack of reruns on newer devices

CPU cooler charts

• Updated with new platforms deployed in 2025, including:
  • 9800 X3D / 9950 X3D
• Uses a more precise acoustic environment:
  • hemi-anechoic chamber
  • Claimed to improve precision vs older open-room methods

Example thermal/noise descriptions include:

• A 157W heat load and noise-normalized comparisons (temperature range mentioned)
• 9950 X3D charts at 276W, with both full-speed and noise-normalized sets

Cooler benchmarking is described as temporarily paused due to team bandwidth, with the page retaining past results and appending new ones.

---

Website content strategy (economic + technical)

The speaker describes a shift in how the work is financed:

• Historically, reviews/cooling/CPU/GPU content subsidized investigations.
• Now, investigations are subsidizing hardware/case reviews because:
  • the market is worse for views and monetization.

Monetization alternatives mentioned

• Keep the site ad-light
  • no third-party ads
  • uses AdSense + limited sponsorship
• Use Patreon and/or store sales to fund free content

Format choice to reduce easy scraping

The speaker notes that converting videos to searchable plain-text articles is better for humans, but is also “super valuable to language models.”

As a result, the channel sometimes prefers PNG/visual charts over text/tables to reduce straightforward scraping/extraction.

---

Planned future directions

Hints for expansion include:

• More features tied to collected data
  • possibly hosting acoustic audio files so users can play test noise samples
• More documentation/blog-style conversions and deeper investigations
• Future areas suggested around “spy devices”
• A forthcoming deep dive involving cameras/microphones

---

Main speakers / sources (as referenced)

• Primary speaker: “Gamers Nexus” / Gamers Nexus staff
• Collaborators referenced:
  • Jensen (referenced in the context of a Cloudflare bot-traffic tweet; attribution is unclear)
  • Wendell of Level One Techs (used for the poisoning test)
• Companies/entities referenced:
  • Cloudflare
  • OpenAI, Microsoft, Anthropic (as scraping/harvesting sources)