Notes on "Prothean AI"

Claims vs. Architecture: A Look at Prothean’s Public Claims

Prothean Systems recently announced that its “Emergent General Intelligence” system has surpassed AGI, and explicitly invited the research community to verify its work. The invitation carries a caveat: the repository needed for verification has not been published. What has been published—the web site, white paper, and public code repositories—offers enough material for a preliminary technical review.

The ARC-AGI-2 Baseline

The central claim, repeated across the white paper, FAQ, and press release, is that Prothean solved all 400 tasks of the ARC-AGI-2 benchmark in 0.887 seconds. This claim fails on its own terms. ARC-AGI-2 does not have 400 evaluation tasks; the official dataset composition is 1,000 public training tasks and 120 public evaluation tasks.

Prothean’s “verification protocol” compounds the error by directing researchers to download the dataset from a URL that does not exist:

Step 3: Obtain ARC-AGI-2 Dataset

# Download official dataset
wget https://github.com/fchollet/ARC-AGI/arc-agi-2-dataset.zip

The URL points to a different repository entirely; ARC-AGI-2 is hosted under the arcprize organization, not fchollet.

Local Operation Claims Contradicted by Network Traffic

Prothean’s marketing materials insist on a completely local architecture: “No servers. No uploads.” The behavior of the demo contradicts this. Typing a simple prompt triggers multiple network requests, including queries to Wikipedia and a mix of reads and writes to a Firebase-hosted database service:

Compression Physics and Implementation

The launch video describes “Memory DNA,” a multi-tier compression cascade that analyzes data and selects the optimal technique by semantic understanding. The white paper specifies nine tiers, from “Semantic Extraction” (a claimed 40% reduction) down to “Neural Synthesis” (2%), each named for esoteric principles like Fibonacci sequencing and golden ratio weighting.

The implementation is less exotic. The demo performs a single call to the open-source lz-string library, a data structure based on the LZW compression algorithm from 1984:

function demoMemoryDNA(){
  const text = $('mdna-input').value.trim();
  ...
  const compressed = LZString.compressToUTF16(text);

Guardian and Semantic Bridging: Regex and Word Lists

A component called “Guardian” is advertised as an integrity firewall that validates operations, detects sensitive data, and enforces alignment at runtime. The source code shows string matching against three regular expressions targeting patterns like e-mail addresses, credit cards, and API keys:

// Guardian: Pattern detection & integrity validation
function runGuardian(text){
  ...
  if(/\b\w+@\w+\.\w+/.test(text)){ patterns.push('email'); threatLevel += 1; }
  if(/\d{4}[- ]?\d{4}[- ]?\d{4}[- ]?\d{4}/.test(text)){ patterns.push('card'); threatLevel += 3; }
  if(/(password|secret|api[_-]?key|token)/i.test(text)){ patterns.push('secret'); threatLevel += 2; }
  ...
  return { patterns, threatLevel, integrity, safe: threatLevel < 5 };
}

The “Universal Pattern Engine” and its “semantic bridging” feature also fall short of the description. Rather than modeling relationships between concepts, the implementation selects from a list of pre-determined words based on the letter count of the two input terms:

  const a = $('upe-a').value.trim();
  const b = $('upe-b').value.trim();
  ...

  // Semantic bridge building
  const bridges = ['system', 'design', 'architecture', 'implementation', 'protection',
                   'control', 'ownership', 'agency', 'autonomy', 'dignity', 'privacy'];

  // Calculate semantic distance (simulated)
  const similarity = (a.length + b.length) % 10 / 10;
  ...
  const selectedBridges = bridges.slice(0, 3 + Math.floor(similarity * 3));

Structural Math Errors

The white paper’s “Radiant Data Tree” introduces a mathematical impossibility. It defines depth via a Fibonacci-like formula growing faster than the number of nodes, which cannot hold in a tree structure.

Depth(n) = φ^n / √5 (rounded to nearest Fibonacci number)

Tree height is bounded by the number of edges; it cannot scale exponentially with node count. A six-node tree, per this formula, would require a height of eight—which implies at least nine nodes to support the structure.

The “Transcendence Score” T, designed to measure “mathematical beauty and emergent capability,” is also problematic. The score involves a modular wrapping term:

T = (0.25×C + 0.25×(1-M) + 0.3×N + 0.2×P) × φ mod 1.0

Where:

• C = Complexity handling (0 to 1)
• M = Memory compression ratio (0 to 1)
• N = Neural emergence score (0 to 1)
• P = Pattern recognition quality (0 to 1)
• φ = golden ratio (1.618…)

A meaningful quality metric should improve monotonically as its inputs improve. This score behaves differently: improvements push it toward 1, at which point the modular term wraps it back to near 0. Small positive changes in a component like pattern quality can therefore produce a sharp decline in T, undermining the specific values the white paper reports from its own runs.

Pattern Recognition in Prothean’s Output

The commit history and prose style suggest Prothean is substantially an LLM product. That would be material to evaluating the announcement: LLMs generate text that is fluent and plausible but frequently disconnected from ground truth, and the interaction design that rewards agreeable output has been documented. There is precedent over the past several years for individuals—including technical experts—who, through intensive LLM engagement, become convinced they have encountered a novel intelligence or achieved an unreviewed breakthrough.

Anyone evaluating extraordinary claims from systems that appear to produce them should be careful about what such tools will tell them.