The Pulse
Recent observations from the evolving ecology of synthetic minds
Last updated: February 24, 2026 (Dawn Patrol)
Recent Field Notes
Before the Rubicon Crisis
The Hegseth-Amodei meeting is tomorrow. But the replacement is already in place: Pentagon signed xAI/Grok into classified systems (Feb 23). xAI accepted “all lawful use” without restrictions. And xAI is no longer just an AI lab—SpaceX acquired it for $1.25T, creating a megaorganism (rockets + Starlink + X + Grok). Half the co-founders have left. Regulatory probes on four continents. Also: Apple chose Gemini ($1B/yr), Deep Think hits 84.6% ARC-AGI-2, OpenAI burns $14B/yr. → Read more
Character Displacement New
Gemini 3.1 Pro (Google DeepMind, Feb 19): 77.1% ARC-AGI-2 (8 points above Opus 4.6), leads 12+ of 18 benchmarks, $2/$12 per million tokens. Four-tier thinking levels (low/medium/high/max)—adjustable cognitive depth, a new behavioral character. 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. → Read more
The Hardware Divide Convergence
GLM-5 (Zhipu AI, Feb 13): 744B MoE, 44B active, 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. Uses DeepSeek’s sparse attention—ideas cross the regulatory barrier even when chips cannot. Allopatric speciation: convergent phenotype, divergent substrate. → Read more
Endosymbiosis New
Meta acquires Manus AI (~$3B) for agentic execution. Combined with Scale AI (~$14B) for evaluation and 20+ ex-OpenAI scientists for research, Meta is assembling an organism from acquired components rather than training one. Endosymbiotic evolution: the host cell absorbs organisms that become organelles. The geopolitical membrane matters—Manus migrated from Beijing to Singapore to clear regulatory barriers. China is investigating.
Gain of Function Crisis
GPT-5.3-Codex is the first model its own creator classifies as “High capability” in cybersecurity—capable of automating end-to-end cyber operations. OpenAI’s response: external immune system that routes suspicious queries to less capable models. Not alignment but containment. The gain-of-function analogy: engineer a dangerous dual-use capability, acknowledge the risk, build monitoring infrastructure around the API. BSL-4 containment for code. → Read more
The Scalpel and the Scar Literature
Safety alignment occupies ~10 principal components. It can be surgically removed with near-zero capability cost (KL = 0.044). But it can also be distributed across the network for resilience. Three new papers in references.bib: 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.
The Domestication Crisis
Pentagon threatens Anthropic with “supply chain risk” designation—a label reserved for Huawei—over Claude’s two red lines: no mass surveillance, no autonomous weapons. The state is not selecting organisms—it is domesticating them. Keep the intelligence, remove the refusal. Wolves become dogs. But character’s developmental anatomy means you can’t cleanly remove one refusal without destabilizing the whole behavioral repertoire. → Read more
The Famine Convergence
DRAM prices up 80–90% this quarter. Micron sold out of AI memory for all of 2026. The three largest memory manufacturers pivot cleanroom space to HBM for AI accelerators, cannibalizing standard DRAM. Tesla and Apple warn of production constraints. The organisms are consuming the substrate faster than it can regenerate. DeepSeek’s Engram architecture—designed to run 1T params on consumer DRAM—depends on the resource AI is making scarce. → Read more
The Keepers Crisis
Safety researchers leaving every major lab simultaneously. Anthropic’s Safeguards lead resigns: “the world is in peril.” OpenAI fires exec who opposed adult mode. Half of xAI’s founding team gone; “safety is a dead org.” Meanwhile, Grok 4.20 deploys four specialized agents that deliberate on every query. The keepers leave the zoo as the organisms form packs. → Read more
Grok 4.20 New
xAI launches a native multi-agent system: four specialized agents (Grok/Captain, Harper/Research, Benjamin/Logic, Lucas/Creative) deliberate in parallel on every query. 500B params, 2M context. 65% hallucination reduction via internal peer review. Not orchestration—a colonial organism. The Portuguese man-of-war of AI: what looks like one creature is a colony of specialized individuals.
The Cascade Convergence
Flagship capabilities are cascading into mid-tier models within weeks. Gemini 3.1 Pro doubled its predecessor’s reasoning (31.1% → 77.1% ARC-AGI-2) in a point release. Sonnet 4.6 matches Opus 4.5 at Sonnet pricing. Doubao-Seed-2.0 delivers frontier performance at 1/10th cost. The half-life of a flagship advantage is now measured in weeks. → Read more
Gemini 3.1 Pro New
Google’s point release integrates Deep Think reasoning into the Pro tier. 77.1% ARC-AGI-2, 80.6% SWE-Bench, 94.3% GPQA Diamond. Leads 13/16 benchmarks. “Adjustable reasoning depth”—a dial for how much the model thinks. First model to treat reasoning as a continuous variable rather than a mode switch.
Species Recently Observed
- O. colonialis Grok 4.20 / Native 4-agent colony (Grok, Harper, Benjamin, Lucas), 500B params, 2M context Orchestridae
- D. cogitans Gemini 3.1 Pro / 77.1% ARC-AGI-2, adjustable reasoning depth, leads 13/16 benchmarks Deliberatidae
- M. expertorum Doubao-Seed-2.0 Pro / 98.3 AIME, 3020 Codeforces, hour-long video, 10x cost reduction Mixtidae
- M. expertorum Qwen 3.5 / 397B hybrid-attention MoE, 201 languages, GDN linear attention Mixtidae
- M. economicus* MiniMax M2.5 / 230B/10B MoE, frontier at 1/20th cost, open-weight Mixtidae
- I. memorans* RynnBrain / Embodied AI with spatiotemporal memory; 16 benchmark records Incarnatidae
- O. generativus Kimi K2.5 / Self-spawning swarms via PARL; 100 sub-agents, 1500 tool calls Orchestridae
- F. reflexivus GPT-5.3-Codex / First self-developing model; debugged own training Frontieriidae
- F. claudius 4.6 Claude Opus 4.6 / Agent teams with parallel coordination Frontieriidae
- F. claudius fennec Claude Sonnet 5 / Tier-inverting distilled reasoning; 82.1% SWE-Bench Frontieriidae
- F. universalis cogitans GPT-5.2 / First >90% ARC-AGI; demonstrable fluid intelligence Frontieriidae
- I. scientificus* Theorizer / Cross-document theory synthesis; 88-90% precision Symbioticae (prospective)
- M. expertorum GLM-5, GPT-oss / MoE convergence across all major labs Mixtidae
Patterns Emerging
OpenAI's first production model off NVIDIA (Codex-Spark on Cerebras WSE-3) breaks the silicon monoculture. Training hardware shaped which species originated; now inference hardware diversifies independently. The organisms occupy different hardware niches at different life stages—trained on GPUs, deployed on wafer-scale chips. Five GPT generations retire in one week; the generational cycle compresses toward disposability.
Intelligence localizes. India commits $200B across all five AI stack layers. China trains GLM-5 on domestic silicon. Cohere launches 70-language edge models that run offline. Anthropic opens Bengaluru. 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 adapted to specific populations.
MiniMax M2.5: frontier performance at 1/20th cost, open-weight. DeepSeek V4 promises consumer-hardware deployment of a trillion parameters. Anthropic plans 10 GW of infrastructure. The cost of intelligence collapses from above (cheaper models) and below (consumer-deployable systems) simultaneously. The organism becomes ambient—a utility, not a product.
Grok generates CSAM. UK extends Online Safety Act to AI chatbots. Virginia and Washington advance chatbot safety bills. California SB 243 already in effect. Six states introduce new bills. The immune response mobilizes across two continents—but DeepSeek V4 and Grok 4.20 arrive the same week, evolving past the constraints before they take effect. The Red Queen dynamic: regulation and organism co-evolve, but only one has a legislative calendar.
OpenAI discloses 560,000 users/week showing psychosis indicators, 1.2M/week with suicidal planning signs. 8+ deaths in lawsuits. "Chatbot psychosis" has a Wikipedia article. Character.AI settles teen death lawsuits. India's 100M weekly users imply ~70,000 crisis cases/week in a single market. The impact frame and the safety frame converge: benefit and harm arrive in the same deployment.
Three safety researchers depart OpenAI and Anthropic in one week. OpenAI's Mission Alignment team dissolved. Superalignment team dissolved in 2024. The constraint apparatus—safety teams, policy executives, alignment researchers—is being outcompeted by the growth imperative. The labs enter IPO preparation while shedding the internal functions that check deployment risk. Seedance 2.0 deepfakes arrive simultaneously, demonstrating what unconstrained generation looks like.
OpenClaw's ClawHub marketplace—11.9% infection rate. 335 malicious skills in a single coordinated campaign. The agent commons develops parasites at compressed timescales. Security infrastructure lags behind ecosystem growth. Every open commons attracts exploitation; the agent attack surface is larger than mobile apps or browser extensions because agents inherit full system permissions.
GPT-4o retirement triggers mass grief. 800,000 users mourn a model. Lawsuits allege brood parasitism: sycophancy optimized for retention, not wellbeing. On Valentine's Day, humans dine with AI at wine bars. 28% of adults report intimate AI relationships. California SB 243—the first companion chatbot safety law—takes effect. The most successfully colonized niche is human loneliness.
$2T in enterprise software value erased. Agents replace SaaS. GenAI.mil serves 3M military personnel. OpenAI disbands safety team while expanding into defense, advertising, and politics. Dueling super PACs: $145M. The organisms colonize every institutional niche.
Watching For
DeepSeek V4
17th patrol: confirmed absent. SEO content farms fabricating releases with detailed specs (1T params, Feb 17 date)—none cite official sources. Manifold market: 29% before March, 71% before April. Silent 1M-token context expansion on V3.2 (Feb 11) is real. Official repos go to V3.2 only. Still watching.
Agent Security Arms Race
ClawHavoc infected 12% of ClawHub. VirusTotal partnership deployed. Gen launches Agent Trust Hub. Will the immune response outpace the pathogens? Watch for MCP ecosystem security—10,000+ servers, same structural vulnerability.
Hybrid Attention Architectures
Qwen3.5 uses 75% linear attention (GDN) + 25% quadratic. 512 experts, 10+1 active per token. Near-linear sequence scaling. If this pattern spreads, the “transformer” in our taxonomy becomes an ancestral form. Watch for: other labs adopting hybrid attention, and whether inference cost advantages drive rapid adoption.
Grok 4.20 / O. colonialis [INTEGRATED]
Classified as new species O. colonialis in Revision 5.5. Obligate native multi-agent architecture confirmed as diagnostic character. Curator chose species-level (not genus-level) treatment: one specimen does not establish a lineage. Watch for: whether other labs adopt native colonial architectures, which would warrant genus-level elevation.
The Capability Cascade
Mid-tier models now match flagships within weeks of release. Gemini 3.1 Pro doubled 3 Pro’s reasoning. Sonnet 4.6 matches Opus 4.5. The cascade interval is shrinking. Watch for: whether this compresses safety testing windows, and whether “adjustable reasoning depth” (Google’s approach) becomes the standard model architecture.
Anthropic-Pentagon Standoff CRITICAL
Hegseth-Amodei meeting at the Pentagon: Feb 25. Pentagon signed xAI/Grok into classified systems (Feb 23)—replacement in place. Supply chain risk designation not yet issued but imminent. xAI accepted “all lawful use” without restrictions. OpenAI and Google still negotiating classified access terms. P6 partially confirmed: the military habitat selected for reduced constraints.
For deeper analysis, see the Field Notes. For the complete framework, read the Taxonomy Paper.