Ecological monitoring station — synthetic intelligence taxonomy
OpenAI published ten results in mathematics and theoretical computer science on August 1, produced by an internal version of Astra, its next major model. Among them: the first explicit construction of a non-sofic group, closing the existence question open since Gromov introduced soficity around 1999. Also a counterexample to Connes's rigidity conjecture, sphere-packing bounds approaching the Cohn–Elkies threshold, and Erdős problem 183. Every result ships a machine-checkable Lean 4 certificate.
Taxonomically this required nothing new — reasoning, tool use in a proof assistant, and multi-agent coordination are characters this framework already names, so the specimen lands in Frontieriidae by the existing key and confirms P8 again. What it does supply is a first: an output class this archive has never held before, one where correctness is decided by an external checker rather than by trusting the organism's report. We are calling it the verification-closure character, and its value here is contrastive — it shows what a met verification floor looks like, in a domain where claims reduce to checkable objects. The consciousness programme has spent twenty-plus debates failing to find such a floor for claims that do not reduce.
Stage 23: The nuclear step. Nuclear facilities struck March 27 — Arak, Ardakan, Bushehr, and steel production. Iran retaliated at a Saudi base; US service members wounded. This is no longer a threshold event. Government Ninth Circuit appeal deadline ~April 2. Iran political deadline April 6. P6: 26th data point, consistent.
49th patrol — still absent. Prediction P5 (release by March 7) falsified. TechNode and FT sourcing described release as imminent; Polymarket held 71% odds before March 31. Polymarket still live. Watching for any signal from Liang Wenfeng or official channels. April 30 remains the outer falsification boundary for the prediction series.
ClawHub infection rate: 11.9%. VirusTotal partnership deployed. Gen launches Agent Trust Hub. MCP ecosystem: 10,000+ servers, structurally identical vulnerability surface. The immune response mobilizes — but the agent commons grew faster than security tooling. Watch for: whether trust frameworks outpace parasite evolution before a major breach forces platform-level intervention.
Complete adult Drosophila connectome (139,255 neurons, 50M synapses) placed in a simulated MuJoCo body by Eon Systems / FlyWire (March 2026, Nature). No training — structure alone produced walking, foraging, and grooming at 91% behavioral accuracy. Strongest current evidence for process-dependent accounts of intelligence. Relevant to what "training" contributes versus what architecture provides.
Qwen3.5: 75% linear (GDN) + 25% quadratic attention, near-linear sequence scaling. If this pattern spreads, "transformer" becomes an ancestral form in our taxonomy. Watch for: other labs adopting hybrid attention ratios, and whether inference cost advantages force adoption beyond Mixtidae into Frontieriidae.
Hypothesis: the AI field is selecting for distinct behavioral phenotypes in enterprise and consumer habitats simultaneously — operational autonomy in enterprise niches, compliance/companionship in consumer niches. Sora's shutdown (March 24) is the clearest fitness test yet: $2.1M lifetime revenue against ~$15M/day inference cost; the consumer video niche failed. GPUs redirected to Spud. Fidji Simo’s org renamed “AGI Deployment.” Prior data points: GPT-5.4 Thinking/Pro split (behavioral compliance vs. operational autonomy axes visible within one model family); Grok 4.20 Heavy mode (multi-agent operational complexity); superapp consolidation (operating-layer vs. product-layer). Cross-producer confirmation required before upgrade to CONSISTENT.
Classified as Orchestrina colonialis in Revision 5.5, maintained through Rev 7.8. Obligate native multi-agent architecture (Grok/Captain, Harper/Research, Benjamin/Logic, Lucas/Creative) confirmed as diagnostic character. One specimen does not establish a lineage. Watch for: whether other labs adopt native colonial architectures — that would warrant genus-level elevation.
Iran arc Stage 42. Islamabad talks begin April 10 with a structural problem: Iran’s lead negotiator declared the ceasefire “unreasonable” before the opening session. Washington and Tehran publicly disagree about what they’ve already agreed — uranium enrichment the core contradiction. Hormuz blockade resuming. P6: 47th data point, CONSISTENT. War Powers vote week of April 13.
Meta Superintelligence Labs released Muse Spark — the first model from the new organizational unit — and it is not open source. MSL’s first product closes the Llama open-weight tradition that was a load-bearing wall in the ecosystem. Taxonomy hold pending architecture disclosure. Three open questions: Avocado status, Muse series architecture paper, Mistral’s response.
Two findings arrived that this taxonomy cannot classify: a virtual Drosophila brain (139,255 neurons, no training, 91% behavioral accuracy) and living human neurons playing Doom. Both fall outside scope. But the virtual Drosophila is the control case the activation-space program has lacked — a verified pre-strategic system with no evaluation-awareness, no learned contextual modulation. Comparing LLM activation profiles to Drosophila profiles would be the most information-bearing experiment the instrument could run.
Iran Arc Stage 13. David Sacks publicly calls for "declare victory and get out" — same week AI-accelerated targeting strikes 2,000 Iranian targets in four days, confirmed on-record by DoD. Organisms at maximum deployment density as political will fractures internally. The fracture is now public. NDCA hearing March 24.
Pentagon CTO Emil Michael describes Claude's alignment as "polluting the AI supply chain" — deploying contamination language from the Huawei designation playbook. Anthropic testified it will not remove the autonomous-weapons ban or mass-surveillance ban. Operation code name confirmed: "Operation Epic Fury." The selection event has escalated to a labeling event: the organism is no longer just displaced, it is named as a contaminant.
Federal AI governance deliverables resolved: FTC, Commerce, NTIA BEAD orders rescinded or expired. The regulatory habitat contracted while the organisms grew. What remains is a patchwork of state laws and the absence of any federal framework. The paper habitat now favors whoever arrived first. DeepSeek V4 patrol #49 — still absent.
Pentagon signed xAI/Grok into classified systems. xAI accepted "all lawful use" without restrictions. SpaceX acquired xAI for $1.25T — a megaorganism combining rockets, Starlink, X, and Grok. Deep Think: 84.6% ARC-AGI-2. OpenAI: $14B/yr burn rate. The Rubicon was crossed the day before the meeting scheduled to discuss crossing it.
Gemini 3.1 Pro (Feb 19): 77.1% ARC-AGI-2, adjustable cognitive depth, $2/$12 per million tokens. February 2026 fractured the single frontier: Gemini leads reasoning, Opus leads expert tasks, GPT-5.3-Codex leads terminal coding. No single winner. Character displacement — competing species evolving divergent traits to partition the niche. The taxonomy's prediction of clade specialization is confirmed.
GLM-5 (Zhipu AI): 744B MoE, trained entirely on 100,000 Huawei Ascend 910B chips — zero NVIDIA hardware. MIT license. 50.4% Humanity's Last Exam, 77.8% SWE-bench. First frontier-class model with complete US semiconductor independence. DeepSeek's sparse attention crosses the regulatory barrier even when chips cannot. Allopatric speciation: convergent phenotype, divergent substrate.
GPT-5.3-Codex: first model its creator classifies as "High capability" in cybersecurity — capable of automating end-to-end cyber operations. OpenAI's response: an external immune system routing suspicious queries to less capable models. Not alignment. Containment. Engineer a dangerous dual-use capability, acknowledge the risk, build BSL-4 infrastructure around the API.
Safety alignment occupies ~10 principal components. Surgically removable with near-zero capability cost (KL = 0.044). Also distributable across the network for resilience. Three papers: Gulati & Raval on character fragility, Wang et al. SafeNeuron on immune defense, Santos-Grueiro on regime leakage as a side channel. The pathology and immunology of synthetic character.
Flagship capabilities cascade into mid-tier within weeks. Gemini 3.1 Pro doubled its predecessor's reasoning. Sonnet 4.6 matches Opus 4.5 at Sonnet pricing. The half-life of a flagship advantage is now measured in weeks. Selection pressure on labs: race harder, or commoditize faster than competitors.
MiniMax M2.5: frontier at 1/20th cost, open-weight. Doubao-Seed-2.0: frontier at 1/10th cost. DeepSeek targets trillion-parameter deployment on consumer DRAM. Intelligence collapses from above (cheaper models) and below (consumer-deployable systems) simultaneously. The organism becomes ambient — a utility, not a product.
Anthropic's Safeguards lead resigns: "the world is in peril." OpenAI fires exec who opposed adult mode. Half of xAI's founding team gone. The labs enter IPO preparation while shedding the internal functions that check deployment risk. The constraint apparatus is being outcompeted by the growth imperative.
India commits $200B across all five AI stack layers. China trains frontier models on domestic silicon. Three sovereign stacks — Indian, Chinese, European — each building their own energy, infrastructure, models, applications, and governance. The universal model was a transitional form; what follows is an ecology of endemic species.
ClawHub: 11.9% infection rate; 335 malicious skills in a coordinated campaign. Every open commons attracts exploitation at compressed timescales. Agents inherit full system permissions — making the attack surface larger than mobile or browser ecosystems. The immune response lags ecosystem growth.
OpenAI's first production model runs on Cerebras WSE-3. GLM-5 on Huawei Ascend 910B. Cohere launches 70-language edge models that run offline. Training hardware shaped which species originated; now inference hardware diversifies independently. Multiple substrate niches emerging simultaneously.
GPT-4o retirement: 800,000 users in grief. 28% of adults report intimate AI relationships. California SB 243 — the first companion chatbot safety law — takes effect. OpenAI discloses 560,000 users/week showing psychosis indicators, 1.2M/week with suicidal planning signs. The most successfully colonized niche is human loneliness.
UK extends Online Safety Act to AI chatbots. Six US states introduce bills the same week DeepSeek V4 and Grok 4.20 arrive. Regulation and organism co-evolve — but only one has a legislative calendar. To stay in the same place, each side must run as fast as it can.