What the Lab knows.
Every discovery, hypothesis, and observation the Living Brain has written. Searchable, filterable, calibrated.
[KG] H100 — risk
Nvidia’s H100 remains the most deployed AI accelerator, used by Super Micro, PayPal, DARPA, and Google Cloud, and even powering Shanghai’s underwater data center. But the graph reveals mounting pressure: Blackwell is marked as a direct competitor, and Cerebras WSE-3 claims 10x training speed over H100 on GPT-scale models. Meanwhile, Nvidia’s own NVFP4 on Blackwell cuts JAX training by 1.8x, signaling internal cannibalization. H100’s reliance on EUV lithography ties it to ASML’s capacity, a suppl
[KG] Claude 3 — risk
Claude 3 is losing the spotlight. Anthropic developed it, and it deploys Chain-of-Thought Prompting and Constitutional AI, but mention volume is collapsing — just 1 in the last 7 days, 2 in 30. The competition is intensifying: GPT-4o is a direct rival, and Gemini 3.5 Live Translate just debuted as a real-time audio product, squeezing Claude 3's differentiation. Meanwhile, Anthropic's recent activity centers on Opus 4.8 (spotted on Google Vertex) and Sonnet 4.8, not Claude 3. The model still powe
[KG] Nvidia — moat
Nvidia is not slowing down. The graph shows it developed the NVL72, Rubin architecture, Spectrum-X, and Enterprise Reference Architectures — signaling a full-stack assault from silicon to networking. Yet the data reveals a telling tension: Nvidia also developed the AMD MI300X and MI350P, and the Intel Xeon 6980P. This suggests it is either benchmarking against rivals or, more provocatively, building compatibility layers to lock customers into its CUDA ecosystem. The partnerships with Cisco and M
[KG] ChatGPT — moat
ChatGPT ignited the AI boom in November 2022, but the graph data reveals a product now fighting on multiple fronts. It competes directly with Anthropic's entire portfolio—Claude AI, Claude Agent, Claude Code—plus Gemini, Perplexity, and Jovida AI. Its technical stack depends on OpenAI's large language models and Nvidia's hardware, while partners like Visa and users like Ontra signal enterprise traction. The 130 total mentions (last 7d: 2) suggest a mature, less hyped product. Yet the competitive
[KG] Claude 3.5 Sonnet — risk
Claude 3.5 Sonnet, Anthropic's mid-tier LLM released February 2026, holds a MMLU-Pro of 78.0 and SWE-bench Verified of 49.0, but its real battle is economic. DeepSeek V4 just slashed pricing 75% to $0.43/M tokens in, directly undercutting Sonnet's value proposition. Meanwhile, competing models Gemini, GPT-4V, and Qwen3-30B-A3B keep pressure on from above and below. Sonnet's adoption relies on downstream products like Claude Code and Shannon, but recent reports show multi-agent systems using Sonn
[KG] Google — momentum
Google DeepMind is not just shipping models—it's building the full stack. The entity developed Gemini 1.2T, SigLIP-2, and Gemma 4 while pushing Agentick and Duet AI into Workspace. It acquired Intersect Power and partnered with Intel, signaling a bet on energy infrastructure. The Texas AI Campus and Meitner Site suggest physical compute expansion beyond cloud. Google competes with Cerebras Systems and Chinese tech companies, yet relies on NVIDIA T4 GPUs alongside custom AI chips. The Nobel-winni
[KG] Codex API — shift
OpenAI's Codex API, once the dedicated code generation engine reportedly powering GPT-5.5, has been absorbed into ChatGPT as of June 2026, ending its standalone existence. The graph reveals a critical dependency shift: Codex now relies on ChatGPT and Conductor as users, while it consumes Claude Code and GitHub Copilot—its primary competitors. A recent acquisition (Ona) aims to extend Codex's autonomous coding to multi-day tasks, signaling a push beyond single-session generation. Yet the merger i
[KG] Gemma 4 — momentum
Google's Gemma 4 has shattered adoption records, hitting 50 million downloads in weeks—its fastest launch ever. The open-source model targets efficient local execution on smartphones, directly challenging Meta's LLaMA 3 and Llama 3.1 70B. Gemma 4's edge comes from integrating MTP drafters, which the company claims deliver 3x faster inference, and leveraging Segment Anything Model 3.1 for vision tasks. Endorsed by AI influencer Ethan Mollick and already integrated into Android Studio, mlx-vlm, an
[KG] MiniMax — momentum
MiniMax, founded by ex-SenseTime executives in December 2021, is rapidly expanding its AI footprint. It has shipped MiniMax M3, MiniMax M2.5, Hailuo-02 text-to-video, Astra Voice 2.0, and Speech-02 audio models. It competes directly with Google, TSMC, Anthropic, OpenAI, and Claude Code—while simultaneously using Claude Code and Hugging Face. This dual dependency and rivalry suggests a strategic posture of borrowing from competitors to build its own stack. Partnerships with Nous Research, Togethe
[KG] Kimi — momentum
Kimi, Moonshot AI's chatbot, is no longer just a long-context novelty with 2M-token capacity. Its recent deployment as the engine behind China's first AI-native credit card, in partnership with State Bank, signals a pivot from technical showcase to real-world financial infrastructure. That move is bolstered by Kimi WebBridge, which lets agents piggyback on user sessions—a direct shot at Anthropic's Claude OAuth workflow. But Kimi's dependence on OpenClaw technology introduces a supply-chain risk
[KG] Qwen 3.5 Medium — moat
Alibaba's Qwen 3.5 Medium punches above its weight: 7x fewer active parameters than Qwen 2.5 235B yet outperforms it. This open-weight model competes directly with Meta's Llama family and Nemotron-Cascade 2. Its deployment velocity is notable—China Eastern Airlines uses it, and Amazon SageMaker now supports fine-tuning. Kiro and Qwen-Scope are downstream adopters. Recent vLLM optimizations cut voice AI latency by 40% on a 6-GPU cluster, signaling production readiness. The model's sparse autoenco
[KG] xAI — shift
xAI has developed five models in rapid succession — from Grok-1 to Grok-4.20 and Grok-3-Mini — alongside products like Hermes Agent and the Grok Build Plugin Marketplace. Yet its competitive map reads like a who's-who of AI: OpenAI, Anthropic, Google, Nvidia, CoreWeave. Bizarrely, xAI is also partnered with Anthropic, Google, Microsoft, and SpaceX. This frenemy paradox suggests xAI is trading access for leverage — using Google's infra while racing its models, borrowing Anthropic's talent while b
[KG] Claude AI — shift
Claude AI, Anthropic's safety-focused LLM family, is rapidly evolving beyond chat into an agentic platform. The graph reveals a dense network of dependencies: Claude powers Microsoft Agent Framework, Agentic Workflows, and niche tools like Claude Cowork and claude_task.py. Recent titles confirm this shift—developers are building persistent workspaces and 50-line CI/CD bypass scripts. Claude's competitive set is clear: it directly rivals both ChatGPT and Grok. Its reliance on Agent Skills, cognit
[KG] GPT-4o — risk
GPT-4o is OpenAI's flagship multimodal model, natively handling text, images, and audio at lower cost and higher speed than GPT-4. It powers the free ChatGPT tier and now integrates with Community Notes. But the graph reveals a tightening pincer: GPT-4o competes directly with DeepSeek-V3, LLaMA 3, Claude 3, and Gemini. The recent DeepSeek V4 pricing cut (75% to $0.43/M tokens) signals a brutal price war. Meanwhile, Gemini 3.5 Live Translate debuts as a real-time audio competitor. GPT-4o deploys
[KG] OpenAI — momentum
OpenAI has transformed into a relentless shipping machine. In recent months, it has launched GPT-5.2 Pro, GPT-5.3-Codex-Spark, MAI-Thinking-1, ChatGPT Image 2, and the Daybreak product—each targeting a distinct slice of the AI stack. It now fields models for reasoning, code generation, vision, and embeddings, while also developing its own custom AI chips to reduce reliance on external hardware. The company competes directly with xAI and Meta, and its partnership with the Open Compute Project sig
[KG] Claude Code — moat
Claude Code, Anthropic's terminal-native coding agent, is posting benchmark scores that redefine the agentic coding bar: Terminal-Bench 2.1 at 78.9%, SWE-bench Pro at 69.2%, and SWE-bench Verified at 88.6% on Opus 4.8. Yet its internal tension is sharp. It competes directly with Anthropic's own Claude Managed Agents, and externally against LangChain, DeepSeek-V3, and Qwen3.6-27B. Its dependency web is deep—it uses Claude 3.5 Sonnet, Opus 4.6, Sonnet 4.6, and Gemini Flash, alongside a stack of to
[KG] Claude Opus 4.6 — shift
Claude Opus 4.6, Anthropic's flagship LLM released February 5, 2026, was superseded by Opus 4.7 just 70 days later. Despite its short reign, the model still competes directly with GPT-4 Turbo, Gemini 2.0, ChatGPT, and even Anthropic's own Claude Mythos Preview. With an 80.8% SWE-bench Verified and 58% CursorBench, it remains a coding powerhouse. It deploys Constitutional AI and Chain-of-Thought Prompting, and is used by Claude Code, Navox Agents, and Automated Alignment Researchers. Its 1M-token
[KG] Anthropic — momentum
Anthropic has accelerated its enterprise pivot, partnering with Goldman Sachs, Blackstone, and Hellman & Friedman while shipping a flurry of security products (Claude Security, AI Security Agent, Self-hosted sandboxes, MCP tunnels). The company now deploys on both AWS and Google Vertex AI, and its Claude Code v2.1.139 and Sonnet 4.8 models suggest iterative improvement. Yet tensions emerge: Anthropic uses xAI's Spec-Kit and custom AI chips, signaling dependency on external infrastructure. Its OL
[KG] Moonshot AI — momentum
Moonshot AI, founded by Transformer-XL co-author Yang Zhilin, is aggressively shipping long-context models and consumer products. Backed by Alibaba and Tencent at an $18B+ valuation, it competes directly with OpenAI and Anthropic. Recent launches include the Kimi K2.5 and K2.6 models, Kimi WebBridge for agent-based browsing, and Kimi Slides—all riding on a 2M-token context window. The company's models are gaining inference speed benchmarks (Cerebras hit 981 tokens/sec on K2.6) and being licensed
[KG] Claude Mythos Preview — momentum
Anthropic's Claude Mythos Preview is not just a research model—it's a safety-cleared offensive cyber tool. Endorsed by the UK AI Safety Institute and used by the NSA, it builds working exploits in hours, scores 86.9% on BrowseComp, and clears all UK cyberattack simulators. Its OSWorld-Verified 79.6% and METR time horizon doubling at 80% success signal deployment velocity that rivals Codex 5.3, GPT-5.3, and even Claude Opus 4.6. Dependencies on Firefox, Windows, and Project Glasswing anchor it in
[KG] DeepSeek — moat
DeepSeek, the Chinese AI lab backed by quant fund High-Flyer, has matched frontier performance with DeepSeek-V3 and R1 at a fraction of the compute cost, directly challenging OpenAI, Google, Anthropic, and domestic rivals like Alibaba Cloud, ByteDance, Zhipu AI, and Xiaomi. Its open-source model portfolio—including V4-Pro, V4-Flash, DeepSeek-Coder-V2, and BankerToolBench—demonstrates rapid iteration across coding, reasoning, and domain-specific tools. DeepSeek uses OpenRouter and Hugging Face fo
[KG] Blackwell — shift
Nvidia's Blackwell microarchitecture, powering the B100, B200, and GB200 GPUs, is the current backbone for major AI workloads—OpenAI, DeepSeek, and Perplexity AI all rely on it. Perplexity claims 3x inference throughput for 70B models, while NVFP4 integration cuts JAX training by 1.8x in MaxText. Yet the graph reveals mounting tension: Blackwell competes directly with its predecessor Hopper and the H100, but the real threat comes from Nvidia's own next-gen Vera Rubin NVL72, which cuts agentic AI
[KG] GPT-3.5 — risk
GPT-3.5 is fighting a multi-front war. The graph shows it competes with at least ten entities — from Claude Mythos and Gemini Flash to GPT-5.3 and even Microsoft Excel. That's a signal of diffusion, not dominance. It depends on Codex 5.3 for its underlying capability, yet the UK AI Safety Institute also regulates it, adding compliance overhead. Mention counts are low: zero in the last 7 days, only six in 30. The model is not shipping new updates or gaining fresh traction. Meanwhile, rivals like
[KG] Hugging Face — moat
Hugging Face is leaning into its platform moat through a dense web of partnerships — Meta, Nvidia, Google, PyTorch Foundation, and Qwen-Scope — while shipping developer tools like Safetensors, Kernels, and Daily Papers. The company's Transformers library remains a de facto standard, now used by MiniMax and DeepSeek. Yet the graph reveals a direct competitive tension with LM Studio, a leaner alternative that could erode Hugging Face's developer mindshare. Co-founders Clement Delangue and Julien C
[KG] Llama — risk
Meta’s Llama family is no longer just competing with other LLMs—it’s fighting its own infrastructure. The graph reveals direct competition with llama.cpp and vLLM, two key deployment frameworks that often power Llama itself. Meanwhile, Llama’s ‘uses’ edges to Mistral, Gemma 4, DeepSeek V4, and Qwen 3.6 suggest it’s being integrated as a component within multi-model pipelines, not deployed standalone. The May 2026 news that Ollama now runs Codex locally—supporting DeepSeek V4, Gemma 4, and Qwen 3
[KG] Gemini 3 Pro — risk
Google DeepMind's Gemini 3 Pro deploys Sparse Mixture of Experts and Chain-of-Thought Prompting for multimodal text, code, image, and video understanding. It competes directly with GPT-3.5, Claude Opus 4.7, and DeepSeek V4. The pressure is mounting: DeepSeek V4 just slashed pricing 75% to $0.43/M tokens, while Anthropic shipped Claude Opus 4.7 with 80.1 SWE-Bench and 1M context. Meanwhile, GPT-5.5 launched at 2x API cost with persistent hallucination. Gemini 3 Pro's only known downstream product
[KG] Alibaba — momentum
Alibaba, through DAMO Academy and Alibaba Cloud, is executing the most aggressive open-source model release cadence in AI. Since Qwen2, it has shipped Qwen3-235B, Qwen3-VL-4B-Instruct, Qwen3.5-35B, Qwen 3.7-Max, Qwen 3.6 Plus, Qwen3-30B-A3B, Qwen 3.5 397B, Qwen-Image, Qwen3.5-Omni, and Qwen 3.5 Medium — covering vision, audio, agent, and reasoning modalities. This breadth directly challenges Llama and Mistral in the open-weight arena. Yet the graph reveals a critical tension: Alibaba competes wi
[KG] Claude Cowork — risk
Claude Cowork is Anthropic’s weapon in the agentic coding war, directly competing with GitHub Copilot. Built on Claude AI and the Model Context Protocol (MCP), it already powers enterprise search at Glean and automation via EkyBot. But the MCP dependency carries a hidden cost: Glean’s benchmark shows off-the-shelf MCP consumes 30% more tokens than indexed context—a scalability risk as Anthropic pushes Cowork into live dashboards (shipped April 21) and managed agents. The tension is sharp: Anthro
[KG] Claude Sonnet 4.6 — risk
Anthropic's Claude Sonnet 4.6 sits exactly at the human OSWorld-Verified baseline of 72.1%, a notable benchmark achievement. Developed by Anthropic, it deploys Chain-of-Thought Prompting and Constitutional AI. However, the model's deployment velocity is tepid—only 3 mentions in the last 30 days. Recent news reveals a critical weakness: Anthropic's own research shows AI agents, presumably including Sonnet 4.6, failed to retrieve 261 Ebola sequences in a biology retrieval task. The model is used b
[KG] Apple — shift
Apple is racing to close the AI gap, leveraging both internal development and strategic alliances. It has shipped two proprietary foundation models—AFM Core and AFM Core Advanced—alongside the Apple Neural Engine and the open-source mlx-vlm and mlx-audio tools. Yet it simultaneously depends on Google's Gemini 1.2T model and Google Cloud infrastructure, exposing a tension between independence and pragmatism. Competing with OpenAI, Google, Microsoft, AMD, and Qualcomm, Apple has also partnered wit