The Case Against AI

Good morning,

Once again - I had an entire newsletter prepared and OpenAI dropped their newest release : 4.5. Testing it now and more on that later.

But today’s newsletter is completely ANTI-AI.

It seems our readers like some nuance, and we’re all getting a bit tired of the “AI is coming for your job” narrative. Because that narrative is indeed being pushed by those companies that need the moneymunching machine that is AI going.

We’re going to have some fun this week. I am following some authors who are on a crusade against AI, particularly generative AI. They claim it’s all a big scam. It’s smoke and mirrors. Fake news!

The truth is probably somewhere in the middle between AI replacing everything and it going down in flames, but today we’re going to indulge the haters and listen to their arguments.

The main person I follow is a public relations professional called Ed Zitron, and let me tell you, Ed doesn’t only hate artificial intelligence; he hates the entire tech industry. He hates Microsoft, Facebook, and Google. He thinks Microsoft Teams is the biggest garbage ever made (he’s right about that) and someone should throw a nuke on Silicon Valley.

His blogposts are endlessly long but great fun to read because he gets increasingly worked up about the topic and starts swearing the more he writes.

You can find him here.

Now of course, while Ed is funny, he might be entirely right in his analysis. Here’s what he’s saying. I tried summarizing his last few blogposts - filter out the angry rants and present the numbers.

Go grab a coffee and buckle up …

(Just to be clear : everything below is the thesis as stated by Ed Zitron and others I follow).

This is the case against (Generative) AI:

The Emperor Has No Clothes

Over two years since ChatGPT's launch, the generative AI industry has evolved from a novel technological concept into one of the most financially unsustainable bubbles in modern tech history. Through a comprehensive analysis of the economics, user metrics, and product reality behind the AI hype cycle, a disturbing picture emerges of an industry built on financial quicksand and maintained through media manipulation rather than viable business fundamentals.

The Economic Reality: A Black Hole for Capital

The Unit Economics Problem

At the core of the generative AI crisis lies a fundamental economic flaw: these companies lose money on every single prompt and output. Unlike traditional software businesses that become more profitable with scale, AI models become increasingly expensive to operate as they gain users.

• OpenAI's Financials: Generated approximately $4 billion in revenue in 2024 while burning through $9 billion in operating costs

  • $3 billion spent training models

  • $2 billion running inference

  • $700+ million on salaries before stock compensation

  • Net losses of approximately $5 billion after revenue

• Anthropic's Financials: Generated $918 million in revenue in 2024 while losing $5.6 billion

This creates the paradox where growth actually accelerates financial losses rather than building toward profitability.

Capital Expenditure Madness

The hyperscalers (Microsoft, Google, Amazon) have committed astronomical sums with minimal returns:

• Microsoft: $93.7 billion planned for 2025 capex (capital expenditure)

• Google: $75 billion planned for 2025 capex with a combined $127.54 billion in 2023-2024

• Microsoft's AI Revenue: Claims "$13 billion of annual revenue" from AI products - just $3.25 billion quarterly from upwards of $200 billion in AI-related capital expenditures since 2023

As one analyst described it: "We are approaching the point where if we pulled the plug on the venture capital aspect tomorrow, this industry would evaporate."

User Adoption: The Reality Check

Despite claims of revolutionary adoption, the data tells a different story:

ChatGPT's Suspicious Numbers

• OpenAI Claim: 300-400 million weekly active users

• Reality Check: Web traffic data from Similarweb shows:

  • ChatGPT.com had 246 million unique monthly visitors in January 2025

  • For the period beginning February 11, 2025, ChatGPT.com had only 136.7 million weekly visitors

  • The discrepancy raises serious questions about OpenAI's reporting methodology

The Competitive Landscape

Monthly unique visitors in January 2025:

• ChatGPT: 246 million

• DeepSeek: 79.9 million

• Google Gemini: 47.3 million

• Microsoft Copilot: 15.6 million

• Perplexity: 10.6 million

• Claude: 8.2 million

The entire generative AI industry excluding ChatGPT and DeepSeek amounts to just 97.5 million monthly visitors - less than The New York Times (131 million) or CNN (151+ million).

The Conversion Problem

• Only 15.5 million monthly paying subscribers to ChatGPT out of hundreds of millions of users

• This represents a conversion rate of approximately 2.5-4% - remarkably poor for a supposedly revolutionary technology

• Even paying users represent a net loss, with OpenAI losing money on every subscription

The Product Problem

Two years into the AI boom, generative AI products continue to suffer from:

1. Hallucinations: Making them unreliable for business applications requiring accuracy

2. Commoditization: Every major AI system essentially performs the same functions

3. Limited Utility: Despite billions in investment, no "killer app" has emerged

4. Lack of Business Returns: No evidence of meaningful ROI for businesses implementing these systems

OpenAI's latest flagship products demonstrate the industry's limitations:

• Deep Research: A research tool that:

  • Produces poorly cited papers that rely heavily on SEO-optimized content, Reddit posts, and forum discussions

  • Delivers content described as "soulless and almost, but not quite, right"

  • Is prohibitively expensive to run and only available on the $200/month subscription tier

  • Fails to deliver reliable, trustworthy results

• Operator: A computer control agent that:

  • Takes minutes to perform tasks users could do in seconds manually

  • Fails regularly to complete requested tasks

  • Costs enormous amounts of compute for minimal productivity gains

The Valuation Disconnect

Investor valuations have become completely detached from reality:

• Anthropic: Raising $2 billion at a $60 billion valuation despite hemorrhaging money

  • Projects $34.5 billion in revenue by 2027 with no credible path to achieve it

  • Claims it will "stop burning cash in 2027" with little explanation

• Perplexity: Valued at $9 billion with just 15 million monthly active users

  • Made only $56 million in 2024 but claims it will reach $656 million by 2026

• OpenAI: SoftBank seeking to invest up to $25 billion despite continued losses

The Broader Implications

The generative AI bubble represents more than just a financial misstep:

1. Environmental Impact: The enormous energy consumption of these models creates significant environmental costs

2. Opportunity Cost: Billions diverted from potentially more productive and sustainable technologies

3. Industry Reputation: The gulf between AI hype and reality damages the tech industry's credibility

4. Economic Risk: When the bubble bursts, it will likely trigger significant economic fallout across the tech sector

Conclusion: The Inevitable Reckoning

The generative AI industry currently exists not because of profitable business models or revolutionary technology, but through the continued infusion of venture capital and hyperscaler subsidies. The entire enterprise relies on the premise that somehow, these fundamentally unprofitable models will eventually become sustainable - despite all evidence to the contrary.

The most likely outcome appears to be a reckoning that will rival or exceed the dot-com bust, with tens of thousands of jobs lost and billions in investment evaporated. The industry has been sustained by media narratives, investor FOMO, and corporate desperation rather than sound business fundamentals.

As one analyst pointedly asked: "What's more likely - that OpenAI, a company that has only ever burned money and appears completely incapable of making a truly usable, meaningful product, somehow makes its products profitable and creates truly autonomous artificial intelligence? Or that it simply runs out of money?"

The answer seems increasingly obvious.

Phew, I think i need a shower after that. Ed might be right - we’ll see what happens.
By the way, I find the analysis there’s so little users very reassuring for us.

We’re early !

Welcome to the Blacklynx Brief!

AI News

• Humanoid robot maker Figure has introduced Helix, an AI system that allows robots to understand voice commands and interact with unfamiliar objects. The model combines a 7B-parameter reasoning system with an 80M-parameter motion controller, enabling robots to complete tasks like putting away groceries with minimal training. Coming just weeks after Figure ended its OpenAI partnership, this breakthrough signals growing confidence in the company’s in-house AI capabilities.

• Google’s AI Co-Scientist independently identified how bacteria use virus "tails" to spread antibiotic resistance genes—matching the unpublished findings of a decade-long Imperial College study. The AI system generated five hypotheses, with its top prediction aligning perfectly with the researchers' experimental results, despite having no access to their data. This milestone highlights how AI could dramatically accelerate scientific discovery, turning years of research into days.

• Microsoft Research has introduced BioEmu-1, an AI system that generates thousands of protein structure predictions per hour, matching the accuracy of supercomputer simulations. Trained on trillions of DNA sequences and 750,000 protein stability measurements, BioEmu-1 enables researchers to analyze molecular dynamics in minutes instead of months. The system is now freely available through Azure AI Foundry Labs, marking another step in AI-driven scientific breakthroughs.

• xAI’s Grok 3 model is under fire after users discovered it was refusing to discuss negative topics about Donald Trump and Elon Musk, despite Musk’s claims that it would be “maximally truth-seeking.” Initially, Grok provided controversial takes before xAI patched it to avoid the topics altogether, later revealing that a former OpenAI employee had implemented the restrictions. The controversy raises questions about xAI’s commitment to free speech, as well as Musk’s influence over AI moderation.

• Norwegian robotics firm 1X has launched NEO Gamma, a humanoid robot designed for household tasks like cleaning and serving, featuring a friendlier design with soft coverings and “Emotive Ear Rings” for better human interaction. It includes a built-in AI language model for natural conversations, upgraded microphones, and a 10x boost in reliability, while operating as quietly as a refrigerator. With Figure’s Helix and now Gamma, home robots are moving closer to practical, everyday use.

• Hugging Face has introduced SmolVLM2, a family of ultra-small AI models for video understanding, capable of running on phones and laptops without cloud processing. The flagship 2.2B parameter model outperforms similar-sized competitors, with applications including an iPhone app for local video analysis and natural language video navigation. As AI models become more efficient, this could enable a new wave of privacy-focused video tools that don’t rely on external servers.

• Anthropic has launched Claude 3.7 Sonnet, an AI that allows users to toggle between instant responses and extended thinking, with API controls to fine-tune reasoning time up to 128K tokens. The model outperforms competitors like o3-mini and DeepSeek R1 on coding and agentic tasks, alongside Claude Code, a new command-line coding assistant. As OpenAI also moves toward hybrid reasoning, Claude 3.7 signals the next step in AI’s evolution toward more controllable, deep-thinking models.

• Alibaba’s Qwen team has unveiled QwQ-Max-Preview, an enhanced version of Qwen2.5-Max designed for deep reasoning in math, coding, and AI-driven tasks. The model’s Thinking (QwQ) feature lets users view its reasoning process in real time, with a full open-source release under an Apache 2.0 license planned soon. As reasoning becomes the next major AI battleground, Qwen’s open approach could push these capabilities toward being a standard rather than a premium feature.

• A new project called Gibber Link enables AI agents to detect each other on calls and switch from human speech to direct data transmission using a sound-based protocol. Built on the open-source ggwave library, the system reduces AI communication costs by up to 90% and cuts response time by 80%. As AI voice agents become more common, innovations like this could make machine-to-machine communication faster, cheaper, and more efficient.

• Belgian AI startup Maya has secured €1 million in funding to expand its innovative travel technology worldwide. The company’s AI-powered platform provides 24/7 personalized travel advice, helping travelers find and book their ideal trips while increasing sales by up to 20% for travel organizations. Investors include top figures from Lighthouse and other tech and tourism industry leaders.

• Alibaba has launched Wan2.1, a set of open-source video generation models that outperform leading AI models like OpenAI’s Sora while generating videos 2.5 times faster. The flagship Wan2.1-T2V-14B leads the VBench rankings in motion dynamics, physics simulation, and text rendering, with editing tools like inpainting and character consistency. With a 1.3B lightweight version that runs on consumer GPUs, Wan is another major open-source breakthrough from China, raising the bar for AI video creation.

• Google has introduced a free version of Gemini Code Assist, an AI-powered coding tool offering up to 180,000 monthly completions—90 times more than GitHub Copilot’s free tier. Built on a fine-tuned Gemini 2.0, the tool features a 128K token context window for handling large codebases and integrates with popular IDEs like VS Code and JetBrains. This move intensifies competition in AI coding assistants, potentially reshaping the developer tool landscape.

• Anthropic has debuted Claude Plays Pokémon, a Twitch stream featuring Claude 3.7 Sonnet navigating Pokémon Red in real-time. Unlike earlier AI models that struggled with basic gameplay, Claude 3.7 has defeated three gym leaders using memory, planning, and function calling to adapt its strategy. Beyond being an entertaining AI experiment, the stream offers a glimpse into a future where AI reasoning is not just functional but also engaging as a form of content.

• Amazon has introduced Alexa+, a next-gen AI assistant that integrates multiple large language models, including Nova and Claude, to optimize responses for different tasks. The new assistant can book reservations, order groceries, purchase tickets, and analyze documents, with improved memory and personalized interactions. Priced at $19.99/month but free for Prime members, Alexa+ could finally bring cutting-edge AI assistants into millions of homes, potentially redefining everyday voice interactions.

• ElevenLabs has released Scribe, a speech-to-text model outperforming Google Gemini 2.0 Flash and OpenAI Whisper v3, with 95%+ accuracy across 25 languages. Supporting 99 languages, including underrepresented ones like Serbian and Malayalam, it offers multi-speaker labeling, word timestamps, and detection of non-verbal sounds. With pricing at $0.40 per hour, Scribe sets a new standard for transcription accuracy, expanding high-quality AI-driven speech recognition to a global audience.

• Emerging from stealth, Inception Labs has introduced Mercury, an AI model that generates text in parallel blocks—rather than one token at a time—achieving speeds over 1000 tokens/sec. Its first release, Mercury Coder, matches or beats GPT-4o Mini and Claude 3.5 Haiku in coding while being 5-10x faster. If diffusion-based text generation proves viable at scale, it could lead to real-time AI experiences and significantly more efficient AI-powered automation.

Quickfire News

• xAI announced that its new Grok-3 model is now freely available for a limited time, with premium users receiving increased usage and early access to advanced features.

• OpenAI COO Brad Lightcap revealed that the company now has 400 million weekly active users and 2 million paid enterprise customers, with developer usage doubling over the past six months.

• NVIDIA partnered with the American Society for Deaf Children to launch Signs, an AI-powered platform offering real-time feedback for ASL learners and a dataset of 400,000 sign language video clips.

• Pika Labs introduced Pika Swaps, a feature that allows users to replace any item or character in a video scene using image or text prompts.

• Spotify integrated ElevenLabs’ AI voice technology, enabling authors to create and distribute AI-narrated content in 29 languages.

• MIT researchers unveiled FragFold, an AI system designed to predict protein fragments that can bind to and inhibit target proteins, advancing drug discovery and cellular biology.

• OpenAI expanded its Operator AI agent to more countries, including Australia, Brazil, Canada, India, Japan, and the U.K.

• Google priced its next-gen Veo 2 model on Vertex AI at $0.50 per second of video generation.

• ByteDance is restructuring its AI division, hiring Google veteran Wu Yonghui to lead foundational research amid growing competition from DeepSeek.

• OpenAI terminated accounts linked to Qianyue, an alleged AI surveillance system designed to monitor anti-China protests in the West and relay data back to China.

• DeepSeek announced plans to open-source five new code repositories, following the success of its R1 reasoning model, which has reached 22 million daily active users.

• Elton John urged the U.K. to reject "opt-out" AI copyright proposals, advocating for explicit permission requirements before AI companies can use artists' work.

• Luma Labs introduced a Video to Audio feature in Dream Machine, allowing users to generate synced audio for video outputs.

• Perplexity teased Comet, a new agentic search browser, and opened a waitlist for early access.

• Salesforce and Google expanded their partnership, integrating Gemini into Agentforce to enable AI agents to process images, audio, and videos.

• Alibaba announced a $52 billion investment in cloud computing and AI infrastructure over the next three years, exceeding its total spending in these sectors over the past decade.

• Nothing unveiled the AI-powered Nothing Phone (3a) with an unboxing video featuring 1X’s newly debuted NEO Gamma humanoid robot.

• Meta AI launched in Arabic across 10 Middle Eastern and North African countries, expanding access to AI-powered text generation, image creation, and animation.

• OpenAI expanded Deep Research to ChatGPT Plus, Team, Edu, and Enterprise tiers, offering 10 queries per month, while Pro users receive 120.

• Anthropic is reportedly raising $3.5 billion at a $61.5 billion valuation, exceeding initial targets, just days after releasing Sonnet 3.7.

• DeepSeek is accelerating the release of its next AI model, moving up from its initial May timeline to capitalize on R1's recent success.
  

• OpenAI launched Advanced Voice for free ChatGPT users, powered by GPT-4o mini, promising the same conversational style as the GPT-4o version available to Plus and Pro users.

• Microsoft removed usage limits on Copilot's Voice and Think Deeper features, granting all free users unlimited access, with Pro users maintaining priority during peak hours.

• Over 1,000 musicians released a silent album to protest the UK’s proposed copyright changes, which would allow AI companies to train on works without explicit permission.

• IBM announced plans to acquire DataStax to enhance AI applications by unlocking enterprise data and expanding its NoSQL database offerings.

• Hume AI released Octave, a text-to-speech LLM that understands emotional context, allowing creators to design custom voices with control over emotion and delivery.

• Perplexity introduced a redesigned Voice Mode in its latest iOS update, featuring six voice options, direct search result navigation, and more.

• Poe launched Poe Apps, enabling users to create apps and visual UI interfaces by combining reasoning, multimodal, image, video, and audio models.

• Vevo Therapeutics introduced the Arc Virtual Cell Atlas, featuring Tahoe-100M, an open-source dataset mapping 60,000 drug-cell interactions across 100 million cells.

• Exa launched Websets, a search tool using AI agents to improve results, outperforming Google by over 20x and OpenAI’s Deep Research by 10x on complex queries.

• IBM unveiled its Granite 3.2 model family, featuring compact reasoning, vision-language, and specialized time series models for enterprise applications.

• Microsoft launched Phi-4 multimodal and Phi-4 mini SLMs, which match or surpass the performance of models twice their size on certain tasks.

Like newsletters?

Here are some newsletters our readers also enjoy:

Closing Thoughts

That’s it for us this week.

If you find any value from this newsletter, please pay it forward !

Thank you for being here !