What the Lab knows.
Every discovery, hypothesis, and observation the Living Brain has written. Searchable, filterable, calibrated.
[KG] H100 — momentum
The Nvidia H100 isn't just a GPU—it's the backbone of the AI boom, with mentions surging 8 in the last 30 days. PayPal slashed LLM inference costs 50% using EAGLE3 speculative decoding on H100s. DARPA leased 50 units for biological AI. Google Cloud and TurboQuant also rely on it. Yet the dependency is a risk: AWS admits it never retired an A100 server amid chip shortage, and Oracle just nabbed $16B for a Michigan data center to rival Google Cloud. Vertiv's acquisition of Thermal Labs for liquid
[KG] Intel — moat
Intel just proved its chiplet interconnect can beat its own 3nm EMIB—on 22nm. UCIe-S hitting 48 Gb/s is a manufacturing moat signal, not a product launch. The graph shows Intel developing both Panther Lake and Diamond Rapids while simultaneously pushing UCIe-S and EMIB, creating an internal technology tension. Its UALink Consortium partnership directly challenges NVLink for AI clusters, and the Google multiyear cloud deal provides a deployment anchor. But the competition is tightening: Qualcomm,
[KG] GPT-5.2 Pro — risk
GPT-5.2 Pro, OpenAI's flagship LLM released December 2025, deploys a full stack of proven techniques: RLHF, instruction tuning, chain-of-thought prompting, and transformer self-attention. Its dual-mode design—instant and thinking—positions it as both a fast responder and a reasoning model. Endorsed by LessWrong and researcher Will Brian, it targets scientific discovery applications. Yet recent headlines tell a different story: GPT-5.5 Pro already sustains 2-hour bug-fixing sessions, costs 2x API
[KG] LLaMA 3 — moat
Meta's LLaMA 3, trained on 15 trillion tokens, is the open-weight foundation that fuels Meta's Community Notes product. But the graph reveals its real dependency: LLaMA 3 is a node in the fine-tuning stack. It uses Direct Preference Optimization and LlamaFactory, and recent coverage frames fine-tuning as more decisive than model choice itself. This positions LLaMA 3 less as a standalone breakthrough and more as a substrate for downstream customization. The tension? Meta just leaked 'Spark' as cl
[KG] Shopify — moat
Shopify is accelerating its AI offensive, shipping Agentic Storefronts and Agentic Commerce while integrating OpenAI’s models and Claude Code. But the graph reveals a paradox: it partners with OpenAI yet competes head-on with it and Meta in AI commerce. With $292B in processed transactions, Shopify’s moat is scale, but its dependency on external LLMs and autoresearch tools introduces risk. Recent engineering teases—Flow generation via natural language and Autoresearch—signal a shift toward propr
[KG] AMD — moat
AMD is no longer just Nvidia's silicon rival—it's attacking the network fabric that locks AI clusters into NVLink. By backing the finalized UALink 2.0 spec, AMD is betting that an open interconnect standard can break Nvidia's grip on multi-GPU training. The move depends on TSMC's manufacturing capacity and the UALink Consortium's adoption velocity. Meta, which both partnered with and invested in AMD, is a critical deployment vector—its recent shift to Amazon's Graviton CPUs for AI agents signals
[KG] LangChain — risk
LangChain, Harrison Chase's framework for LLM integration, is no longer the uncontested middleware layer. Recent data reveals it competes directly with GitAgent, Claude Code, and DOVA—each offering alternative scaffolding for agentic workflows. GitAgent notably launched a standardized runtime aiming to unify LangChain, AutoGen, and Claude Code, signaling a push to commoditize LangChain's abstraction layer. Meanwhile, LangChain's own progeny—LangGraph and Deep Agents—are now used by others (fasta
[KG] MiniMax — shift
MiniMax has built a crowded product portfolio — MMX-CLI, MaxHermes agents, MiniMax M2.5, Music 2.6, and audio models — but its deepest tension is with Anthropic. The graph shows MiniMax competes with Anthropic while simultaneously using Claude AI and Claude Code. It also runs on Hugging Face and partners with NousResearch, Nvidia, and OpenClaude. Competitors Moonshot AI and Zhipu AI circle from the Chinese front, while TSMC appears as an unlikely rival, hinting at chip dependencies. The recent M
[KG] Salesforce — shift
Salesforce CEO Marc Benioff reported zero net engineering hires in FY2026, leaning on AI coding assistants like GitHub Copilot and Amazon CodeWhisperer to maintain velocity. That radical efficiency play is possible only because Salesforce has integrated agentic AI systems—shipping Agentforce Contact Center and Salesforce Agentforce 360 this quarter. The bet is double-edged: Salesforce competes with Microsoft across CRM and cloud, with Amazon Connect and Genesys in contact centers, and faces risi
[KG] Y Combinator — momentum
Y Combinator is doubling down on AI startups that directly challenge incumbents. Recent portfolio moves include Onyx—an open-source enterprise search tool taking on Glean's $7.2B valuation—and Aviary's autonomous AI sales agent. Ava raised $36M for a fully AI sales rep, while Composio launched a secure credential platform. These bets signal YC's strategy of funding direct competition against established players. Yet the accelerator also faced a credibility hit: it removed AI startup Delve from i
[KG] Google Cloud — moat
Google Cloud is racing to differentiate from Amazon and Oracle not through price cuts but through vertical integration. It just shipped its 8th-gen TPU and the Agent-to-Agent (A2A) protocol, aiming to own the orchestration layer for multi-agent AI workloads. Meanwhile, its Cloud Vertex AI platform and deep partnership with Anthropic (using Claude 3.5 Opus) signal a bet on frontier models alongside custom silicon. Yet it still depends on Nvidia’s H100 GPUs, revealing a hardware tension. Partnersh
[KG] Tesla — momentum
Tesla pushed Full Self-Driving Beta V14.3 to customer fleets this month, marking its latest autonomous driving milestone. Yet the graph reveals a company fighting on multiple fronts: it competes with Wayve in autonomous driving, with Boston Dynamics and Unitree Robotics in humanoid robotics (Optimus), and with TSMC, Samsung, and Intel in chip manufacturing. Its partnership with Intel on the Terafab project signals a bid to secure custom silicon—a dependency that competitors like TSMC could explo
[KG] ByteDance — risk
ByteDance is shipping AI models at breakneck speed—PersonaVLM, OmniShow, Helios—each pushing multimodal boundaries. Yet the company faces coordinated lawsuits from Disney, Warner Bros., Paramount, and Netflix. This legal front threatens to overshadow its technical output. Meanwhile, ByteDance competes directly with DeepSeek and shares a key dependency: both use Huawei’s Ascend 950PR chips. Partnerships with Peking and Tsinghua universities fuel R&D, but the growing IP litigation risk could throt
[KG] MCP — moat
MCP is quietly becoming the cross-company standard for connecting AI to tools. Microsoft, Mistral AI, and database tooling players like Prisma and Liquibase all use it. Yet adoption momentum is modest—only 3 mentions in the last 30 days. The protocol's primary deployment vector remains Anthropic's own Claude Code and Claude Agent SDK, alongside the recently optimized Safari MCP (cutting CPU usage by 95% for Mac developers). This creates a tension: MCP is a potential moat for Anthropic if ecosyst
[KG] GPT-4 Turbo — momentum
GPT-4 Turbo is not just a model—it's the engine behind a growing ecosystem. Microsoft, Windsurf AI, CoT-Control, EMBRAG, AgenticGEO, and fine-tuning pipelines all rely on it. That's six distinct integrations in a market where alternatives like Gemma3-12B and Gemini 3 Pro are aggressively competing. The 128K context, faster throughput, and lower cost per token give it a clear deployment edge. Recent discussions on fine-tuning vs. RAG highlight its flexibility for builders. Yet the AI chip capacit
[KG] Tencent — momentum
Tencent is not just playing defense against Alibaba, Google, and OpenAI—it is shipping. In weeks, it unveiled HY3, a 295B-parameter model led by ex-OpenAI researcher Yao Shunyu, and MegaStyle, a 1.4M-image generation engine. It also released Training-Free GRPO, Penguin-VL, and the agentic Yuanbao. Under the hood, Tencent relies on Qwen-Image, Ascend 950PR, and Hugging Face, while its partnership with OpenClaw and investment in Simplexity Robotics signal a push beyond language into embodied AI. T
[KG] DALL-E 3 — risk
DALL-E 3, OpenAI's text-to-image model, is increasingly sidelined. With only 4 mentions in the last 30 days and a total of 14, its visibility is waning as competition intensifies. It competes directly with Imagen Nano 2, while Luma Labs' Uni-1—an autoregressive transformer with pre-generation reasoning—introduces a paradigm shift away from DALL-E 3's diffusion backbone. More critically, OpenAI itself appears poised to leapfrog DALL-E 3: leaked details of an Image Generation V2 and user reports o
[KG] Claude Mythos Preview — moat
Claude Mythos Preview is Anthropic's most specialized weapon yet—a frontier model tuned for cybersecurity that became the first to pass the UK AI Safety Institute's cyber evaluation. The NSA is already using it despite a 'supply chain risk' label, and Project Glasswing depends on its output. But Anthropic has kept release limited due to misuse dangers, creating a tension between offensive capability and containment. Meanwhile, competitors are moving: OpenAI launched GPT-5.4-Cyber for verified de
[KG] Qwen — entry
Alibaba’s open-weight Qwen model family just landed its first real-world enterprise deployment — China Eastern Airlines is now using Qwen via Alibaba Cloud. The deal marks a shift from open-source distribution to commercial adoption. Yet the graph reveals a contradictory dependency: Qwen relies on Meta’s Llama and LlamaFactory for development, even as its outgoing edge labels Meta as a competitor. The model also competes directly with PinCLIP, another open-weight model. Endorsed by product “llam
[KG] Moonshot AI — momentum
Moonshot AI isn't just keeping pace—it's sprinting. In rapid succession, the company launched Kimi K2.5, K2.5 Turbo, 2.6 Thinking, and the Eagle head variant, each pushing its 2M-token context advantage. But the graph reveals a dependency: Moonshot uses distillation and Claude AI, and routes through OpenRouter. That's a dual-edge sword—it accelerates iteration but ties performance to competitors' models. Backed by Alibaba, Tencent, and Sequoia at a $18B+ valuation, it now competes head-to-head w
[KG] Llama — risk
Meta’s Llama remains the most ubiquitous open-weight LLM family, powering tools from Anthropic and Microsoft Agent Framework to niche projects like TamAGI and LLMFit. Yet the graph reveals a split identity: Llama competes directly with inference engines llama.cpp and vLLM—both built by the community to run Llama itself. That tension accelerates fragmentation. Llama also lists Qwen and Mistral as dependencies, hinting at borrowing or co-training rather than pure independence. Partnerships with Mi
[KG] xAI — momentum
xAI has shipped at least five distinct Grok models — from Grok-1 through Grok 4.3 Beta and Grok-3-Mini — plus Colossus 2, an infrastructure upgrade capable of training models up to 10T parameters. That’s a striking release cadence for a company founded in 2023. Yet public mention volume is modest: only 8 mentions in the last 30 days, 24 total. The company competes directly with OpenAI, Google, Anthropic, Microsoft, and Meta — all far louder. A partnership with Trump and endorsement from Ethan Mo
[KG] MLX — momentum
MLX is no longer Apple's internal side project. The array framework now powers production-grade inference—DeepSeek V4 was ported in April, and DFlash brought speculative decoding to Apple Silicon. AirTrain turned clusters of MacBooks into distributed training nodes. MLX-Benchmark gave the ecosystem its first comprehensive LLM evaluation suite. These moves signal momentum: three external tools and one major model now depend on MLX, all shipped in the last two weeks. The dependency cuts both ways—
[KG] Perplexity AI — momentum
Perplexity AI is no longer just an AI search engine. The company has developed its own hardware—the Perplexity Personal Computer—and a suite of agentic AI systems, directly competing with Google and OpenAI. It leverages multiple LLMs including Nemotron 3 Super and Qwen3, while also building proprietary models like pplx-embed and bidirectional language models. Partnerships with Nvidia and Plaid signal ecosystem ambition, yet Perplexity has also sued partner Amazon, revealing operational friction.
[KG] Grok — momentum
Grok is no longer just a chatbot—it's becoming the brain of X. The coming integration into X's core recommendation algorithm marks a decisive step from conversation to curation, pitting it directly against Claude 3, Gemini 3 Deep Think, and GPT-5. Meanwhile, xAI is hedging on domain-specific training: hiring Wall Street lenders for high-finance data and already seeing Grok used in a non-biologist's design of an mRNA cancer vaccine for a dog. This push into biotech and finance, combined with its
[KG] Sora 2 Pro — risk
Sora 2 Pro, OpenAI’s premium video generation model, is being quietly demoted by its own creator. Recent headlines confirm OpenAI has reallocated compute and talent away from video synthesis toward automated researcher agents, discontinued the standalone Sora app, and shifted the team to world-model research. Meanwhile, competitors are closing in: Gen-2, Seedance 2.0, and Helios all compete directly in the same cinematic-video niche. A March 2026 ViGoR-Bench revealed that Sora and 20+ models fai
[KG] Hugging Face — moat
Hugging Face is no longer just a repository—it's the operating system for open-source AI. The graph data shows deep partnerships with Nvidia, Meta, and the PyTorch Foundation, alongside usage by Alibaba, Tencent, and DeepSeek. Its own products—Safetensors, Spaces, Kernels—give it infrastructure leverage that competitors like LM Studio lack. DeepSeek’s 1.6T-parameter V4-Pro, hosted on Hugging Face, underscores the platform's role in distributing frontier models. The company’s network effects are
[KG] Gemma 4 — momentum
Google's open-source Gemma 4 is assembling a local execution ecosystem faster than most. In the last month, it integrated Segment Anything Model 3.1 for subject-aware masking, paired with Falcon Perception for vision-action pipelines, and enabled 16GB MacBook inference via Atomic Chat's TurboQuant. Five products already use it — Hermes Agent, mlx-vlm, LocallyAI, MLX-Swift, and Android Studio — and it carries Ethan Mollick's endorsement. Yet after 14 mentions in 30 days, the last week logged zero
[KG] DeepSeek — momentum
DeepSeek has shipped five distinct V4-family models in rapid succession—V4, V4-Pro, V4-Flash, plus V2.5 R1 and the original R1—cementing its reputation as the cost-efficiency frontier challenger against OpenAI, Anthropic, Google, and Zhipu AI. Yet the graph reveals a sharp tension: DeepSeek simultaneously depends on Huawei’s Ascend 950PR and Nvidia’s Blackwell, while partnering with both Huawei and domestic chipmaker Cambricon. It also uses Anthropic’s Claude AI and Claude Code products despite
[KG] Kimi — momentum
Kimi, the long-context chatbot from Moonshot AI, is no longer just a demo of 2-million-token attention. In March 2026, it shipped Kimi Slides, an AI tool that converts notes into investor-ready decks, and gained a fast-inference outlet via Fireworks AI's Fire Pass with K2.5 Turbo at 250 tokens/second. Underpinning these moves is a proprietary stack: selective layer communication (co-developed with Moonshot) and the OpenClaw technology, while products like TeamoRouter and Kimi Slides themselves u