The Four AI Wars: From GPU Wealth to Useful Work — Part 1 of 3

GPU wealth does not guarantee a business

Inflection had reportedly raised $1.3 billion less than a year earlier, yet much of its team appeared to be moving to Microsoft. Did the move expose the compute requirements of competing at the model frontier? A guest estimates that Inflection had roughly 22,000 H100 GPUs—wealthy by startup standards, but small beside Microsoft. Compute was only part of the problem: Inflection had built its own model and experience, but its interface remained similar to competing products.

A startup with fewer GPUs could instead solve a problem the large labs overlooked. The guest recalls an unnamed enterprise AI lead reporting $10 million saved through translation, though no timeframe, baseline or cost accounting accompanies the anecdote. They consider strategy a plausible explanation for Inflection’s move, while acknowledging that they do not know the reasons behind it.

That distinction gives the Four Wars framework in this Latent Space conversation its practical force. Developed while recapping 2023, it sorts durable AI news by competition over scarce resources, especially talent. Its four fronts are data, GPU-rich versus GPU-poor companies, multimodality, and a fourth category named RAG and operations, whose mechanics are not developed here. Individually plausible company strategies cannot all succeed when they compete for the same resources.

Reported departures at Stability AI reinforce concern about heavily funded model companies without settled business models. One guest anticipates consolidation—a brief cold front rather than an AI winter. The host adds that many less-resourced startups already pursue differentiated applications rather than frontier models. Adoption and consolidation can grow together: enterprise investment and recurring personal use need not prevent attrition among crowded chatbot builders and image generators.

Workflow innovation and the multimodal advantage

Coding assistants show how product innovation can survive incumbent advantages. A guest recalls expectations that GitHub Copilot would dominate, then points to Cursor’s adoption and excitement around Devin. Cursor expanded the interaction from single-line completion to chat and multi-line edits. Those changes opened ways to compete beyond the quality of code completion alone.

The guest sees less room for experimentation in image-generation interfaces, where competition had concentrated on better images rather than better control of the generation process. Sora’s video generation offered a different creative medium. In this interpretation, creative users seek new effects on an audience, while developers often arrive with a particular frustrating task. A new medium or workflow can therefore matter more than another increment in image quality.

The multimodality question is whether companies working across media such as text, images and video gain an advantage over specialists. The host contrasts DALL-E 3 within OpenAI’s portfolio with Midjourney. A respondent says Sora shifted their view toward the benefits of developing several models under one roof: they describe DALL-E 3 using GPT-4 Vision and GPT-4 to bootstrap synthetic data—training material generated by models—and Sora using related methods. Progress in one modality can supply training material for another.

The discussion leaves two routes to a defensible product: - Broad providers can compound capabilities. Several strong models in-house can help generate training data across modalities, giving their developers an advantage over a single-modality startup. - Specialists can bring new people into an activity. Suno’s music generation draws enthusiasm from people who enjoy music but do not identify as musicians. The host treats that engagement as a strong early product signal. The respondent still sees room for music specialists, partly because they do not expect OpenAI to pursue music soon—a judgment from the time of the conversation.

Data licensing raises costs and restricts access

Reddit’s reported roughly $60 million agreement with Google before its public offering makes the data war concrete. A guest emphasizes non-exclusive licensing: Reddit could license the same data to other buyers. They regard vote scores as valuable training information and predict that data will become model companies’ largest expense after GPUs as licensing payments and litigation grow.

For open-source developers, restricted access creates a budget problem: how can builders without tens of millions of dollars for annual licenses obtain good datasets? Synthetic data offers another source, but introduces two separate tests: - Performance: which generated material improves a model, and in what quantity? - Commercial sourcing: is the method used to obtain that material acceptable for its intended commercial use?

Transparency complicates both questions. One participant criticizes Mira Murati’s answer about Sora’s training data in a Wall Street Journal interview and argues that companies have little incentive to disclose their sources. They compare the unsettled situation to the transition from Napster to Spotify. Another agrees that social debate and court decisions will help determine the outcome alongside technical work.

The participants resist treating synthetic data as uniformly good or bad. Reports of degraded models sit alongside tentative accounts of strong models trained with synthetic material, without enough training detail to resolve the comparison. A respondent uses cholesterol as an analogy: type and quantity matter. They predict widespread use and cite AlphaGeometry as an encouraging example, without explaining its method. The practical challenge is selecting useful generated material and avoiding poor sourcing or performance outcomes.

Recurring work becomes the model evaluation

Claude 3 Opus changed one participant’s podcast workflow. Their tool creates chapters, recognizes named entities and summarizes episodes. After finding Claude 2 unusable for that work and relying on GPT-4, they compared Opus with GPT-4 side by side and preferred Opus. Availability shaped the comparison: they had not obtained access to Gemini.

The advantage was chiefly subjective writing quality. After roughly 50 podcast episodes and dozens of generated outputs, the participant had become sensitive to GPT-4’s recurring vocabulary and phrasing. Their firsthand claim of a substantial improvement applies specifically to this podcast task.

Another participant observes people switching recurring tasks to Claude and discussing their actual work rather than benchmark scores. Stable workflows let users compare models directly. Once people know what a useful output looks like, they can select models much as they select ordinary software. The participant interprets these switches as a gain in Anthropic’s credibility more than a decisive setback for OpenAI.

Competition also raises expectations for GPT-5. Long context—the capacity to accept more input at once—and multimodality increasingly look like baseline capabilities to one participant. A new flagship would have to match competitors and offer something more. That competitive argument does not establish the accompanying sweeping claims about million-token inputs or perfect recall, which lack model-specific measurements.

OpenAI’s public posture supports opposing interpretations: one participant sees apparent calm as a sign of a major advance in reserve; another leaves the same recent interview thinking GPT-5 is distant. Reports of GPT-4.5 and agent projects remain rumors. The latter participant treats version numbers as marketing and release decisions imposed on continuous improvement. Their recommendation is to keep building rather than infer readiness from public remarks.

Open availability, local computing and who pays

The participants welcome a market with several credible model providers. Mistral Large is judged impressive for a new startup, though one participant places it below GPT-4 class without presenting an evaluation. Its commercial release raises a funding question: the host perceives community acceptance of selling some models to finance continued open-source development.

Anticipation of Llama 3 exposes a distinction between open availability and practical deployability. One participant relays unverified hallway rumors of a 260–300-billion-parameter model and objects that ordinary users could not run something so large. Another replies that difficulty running a model does not determine whether it can be open source. The potential audience is then reframed as organizations building competitors to OpenAI or Anthropic, rather than people using laptops.

Apple’s MM1 effort signals in-house model training, with a roughly 30-billion-parameter multimodal model discussed here. A participant sees Apple’s larger opportunity in trusted personal computing: increasingly intimate AI inputs could make users reluctant to send everything to cloud services. They regard Apple’s control across chips and devices as an advantage in meeting that demand.

The rumored Apple–Google agreement leaves two opposing payment models open: - Google pays for distribution: default placement for Gemini makes Apple’s user relationship valuable. - Apple pays for capability: useful AI becomes a feature customers expect before buying a device. Who pays would reveal whether the greater bargaining power lies with the entry point or with the model capability that makes the device desirable.

Meta’s stock-market recovery prompts a more personal reflection on Zuckerberg. Participants associate renewed enthusiasm with Meta’s AI emphasis, without establishing that AI caused the gains. The host sees Zuckerberg’s direct-to-camera headset comparison as effective communication even for viewers who disagree. A guest also finds lessons in his personal growth and perseverance, while keeping that admiration separate from judgments about Facebook’s social impact.

Reliable agents need a bounded job

AutoGPT illustrates the gap between an appealing goal and dependable execution. A participant recalls roughly 125,000 GitHub stars in six weeks, followed by about 150,000 a year later. They acknowledge that many interested users may already have starred it, but see enthusiasm plateauing after the promise of handing an agent any goal. Businesses moved toward narrower tasks they could make reliable.

Vertical agents specialize in a domain. Mike Conover’s Brightwave is presented as an example focused on financial research; security, compliance and legal work offer other applications. Narrowing the domain also makes evaluation tractable. A team needs a bounded set of tasks to determine whether an agent succeeds, whereas unrestricted work is difficult to test. Enterprise evaluation tooling remains fragmented in the participant’s account because production use cases are only beginning to become established.

Robotics supplies another setting for agents. An attendee describes robot dogs and industrial robots at Nvidia GTC, with Omniverse providing digital twins—representations of environments used to train robotic agents. For office work, the same participant proposes the full stack employee: someone working across organizational functions through agents and AI-enabled tools. They argue that this broad outcome requires narrower, more reliable applications along the way.

Another participant argues that prompt engineering and agent orchestration cannot deliver general intelligence without substantial model improvements. Even Devin, which they regard as especially advanced, may still suffer failures, latency and high costs; those concerns are speculation rather than hands-on findings. Their product criterion is concrete: does the system reliably remove routine work someone would otherwise do or delegate?

A participant producing AI tutorials sees role-specific solutions as especially useful, even when less technically glamorous. Their own habit of making one or two YouTube thumbnails with Midjourney each day shows how a general tool becomes valuable through repeated use. Broad change can accumulate from many such small improvements.

Vision-driven desktop agents remain a possible general-purpose exception. These systems would watch a screen and operate software visually. A participant associates that promise with Adept and imagines a software layer comparable in significance to a browser or operating system. They describe the approach as not yet mature: domain-specific applications offer the immediate opportunity, while a broadly useful computer operator remains under development.

New architectures could change what models can process

RWKV and Mamba are presented as alternatives to conventional transformers, with Mamba identified as a state-space model. The discussion emphasizes their promise of better scaling rather than explaining their internal mechanics. The motivating thought experiment is vastly greater input capacity: what would change if a system could accept 200,000 hours of video? That is a hypothetical application question, not a reported capability.

A participant reports encouraging RWKV results against Llama 2 at seven billion parameters with comparable training, but leaves the benchmarks unspecified. Results at 7B do not establish performance at 70B. Both RWKV and Mamba are presented as open-source options available to explore at the time.

Diffusion and transformer architectures are also described as converging: Stable Diffusion already incorporated transformers, and Sora uses a diffusion transformer. Stable Diffusion 3, Hourglass Diffusion and SDXL Turbo illustrate continued architectural experimentation. The participant argues that perceptions of slow image generation and poor text rendering lag newer models, though no timing or quality measurements accompany that assessment.

The wildcard is text diffusion: generating chunks of text together rather than one token at a time. A participant reports that Midjourney confirmed exploring the approach that day. From the conversation’s 2024 vantage point, they see possible production use of emerging architectures two to three years ahead.

Wearables acquire personal context—and need social acceptance

Always-on recording requires acceptance beyond the wearer. An enthusiastic wearable user notes that even podcasters do not record every conversation. The participant discloses an investment in Tab, reports buying Rabbit, and describes seeing two Humane devices in San Francisco. Enthusiasm for physical AI experiments coexists with their prediction that Humane will fail.

Hardware preorders also supply financing before delivery. The participant recalls paying roughly $200 for Rabbit and $600 for Tab, with about six months still to wait, and contrasts that willingness with resistance to an $8 monthly software subscription. Open Interpreter’s newly announced $99 push-to-talk microphone, connected to a laptop, offers another tangible AI interface.

A participant relays Avi’s view that consumers prioritize utility over privacy, drawing a comparison with social media, but questions whether AI will encounter stronger resistance. This is an argument about likely behavior, not a measured consumer preference.

The strategic value returns to the data war: wearables could acquire personal context unavailable elsewhere. Everyday conversations may reveal a person’s concerns and train of thought without deliberate note-taking. A participant therefore expects the software and exclusive context to matter more than any particular device shape. The form factor delivers the context; the useful AI experience depends on what the software does with it.

Faster inference could make new applications worthwhile

Groq’s Mixtral demonstration draws attention to generation speed, reportedly reaching 500 tokens per second. Its concurrency and model configuration are unspecified. Chip prices also resist simple comparison: a participant cites approximately $2 million for a Cerebras chip and $40,000 for an H100 while explicitly cautioning that these are not equivalent units of capability.

Serving economics depend on several interacting choices: - On-chip memory versus model growth: placing memory on the chip can help language models, but model sizes may outpace fixed capacity while hardware development takes much longer. - Compute price versus concurrency: Groq’s reported argument is that a higher price per unit of arithmetic work can still produce favorable economics when serving many requests simultaneously. - Upfront capital versus later upgrades: providers may invest more initially and pursue high utilization, or commit less and upgrade later. A chip price or raw compute figure alone cannot settle the serving decision.

Lemurian Labs and MatX are mentioned as smaller contenders working on mathematical infrastructure and execution instructions as well as chips. One participant predicts that within two to three years, comparable serving speeds will let developers care less about underlying hardware again. Another disputes the limited-impact framing: capturing even 1% of Nvidia’s business would make a startup consequential, and users could feel the competition through faster inference.

The latter participant forecasts typical generation speeds rising from roughly 50–100 to 500–2,000 tokens per second during 2024, citing unnamed suppliers without specifying workloads or configurations. The consequential idea is speed as an application capability. An order-of-magnitude improvement could make tedious interactions enjoyable and allow background work to finish unobtrusively, changing which tasks feel worth delegating.

Part 1 ends with the competitive mechanism unresolved. Are accelerator startups pushing incumbents to improve, or benefiting from the same technological advances? Another participant wonders whether Nvidia’s next annual release could surpass competing products arriving that year. Faster AI could expand the application possibilities either way; which suppliers capture the opportunity remains open.