OpenAI : The Weakest Link

Good morning,

In this week’s edition of the Brief we’re pointing our gaze towards OpenAI.

As the biggest AI studio in the world they are responsible for about 80% of headlines. And not always in a good way.

Sure, they’re the studio behind ChatGPT, arguably the most well-known large language model. But, internal strife that should’ve been handled discreetly has exploded onto the frontpages of many newspapers. Think about the entire saga of Sam Altman’s firing or the very public spat between Altman and Elon Musk - one of the original founders of OpenAI.

Sam Altman is a VERY controversial figure in Silicon Valley - very good at generating hype, spinning narratives and looking innocently into a camera. On the other hand his nickname is “Scam Altman” and he’s been called “a snake” - as documented in this newsletter.

We don’t know about all that, but what we do know is that they don’t always seem to know what they’re doing at OpenAI as it’s reflected in the naming convention of ChatGPT. I don’t know about you, but I can’t follow anymore. There’s a 4.5 , an o1, an o3 - o3 mini, 4.0 and now since this week 4.1 - a new model geared towards coders.

Luckily, they seem aware of the issue.

Then this week came the news that OpenAI is planning to start a social network - in order to compete with X.

Consider me confused.

This seems more a move against Elon Musk and an emotional move than a well thought-out strategy.

It smells like desperation.

Then there was the “image generation fiasco”.

Few weeks ago OpenAI debuted their new image generation engine and since then we've been bombarded on social media and newssites even with first "Studio Ghibli"-style images (which is now forbidden as Studio Ghibli has filed a cease and desist) - closely followed by everybody and their uncle creating an action figure of themselves.

The latest trend is to take a picture of your pet and let "ChatGPT" turn it into a human.

While this - granted - provides some comic relief - I can't help but wonder about the WASTE.

Behind every amusing AI-generated image of your cat reimagined as a Victorian gentleman lies a disturbing reality: an industry burning through billions of dollars at an unsustainable rate.

Our friend Ed Zitron, OpenAI and AI hater “par excellence” did a recent deep-dive financial analysis of OpenAI and it paints a troubling picture.

Now Ed REALLY HATES OpenAI - so I’m a bit skeptical but even if he’s half right - it’s downright troubling. I invite you to take the time to read the full thing but it’s loooooong.

It also does not account for the impact that Trump’s tariffs will have.

Despite its astonishing $300 billion valuation and record-breaking $40 billion funding round, OpenAI operates on a fundamentally broken business model, the article claims.

The company spends approximately $2.25 for every $1 it earns. This isn't just slight unprofitability — it's a financial structure requiring perpetual capital infusion at unprecedented scales. OpenAI needs at least $40 billion annually just to maintain operations, with projected expenses of $320 billion over the next five years.

Most concerning is OpenAI's GPU shortage crisis. CEO Sam Altman has repeatedly warned about capacity issues, noting how their image generation features are causing their "GPUs to melt."

While users experience the fun output, the infrastructure struggles to keep pace with demand. Each free user who generates a handful those pet portraits before abandoning the platform represents another incremental loss.

The company's expansion plans rely on partnerships with unproven startups like Crusoe (a former cryptocurrency mining company) and Core Scientific (a recently bankrupt crypto miner) to build massive data centers. These companies have never constructed AI data centers at the scale required, yet OpenAI's future depends on their successful delivery of infrastructure that even tech giants like Microsoft have backed away from building.

Perhaps most alarming is SoftBank's role in this house of cards. To fund OpenAI, SoftBank is stretching its financial limits to dangerous levels, with rating agencies warning that its "financial condition will likely deteriorate" as a result. Yet even with this support, OpenAI faces a critical deadline: convert to a for-profit entity by December 2025 or lose $10 billion in committed funding.

Basically, there is a realistic chance that OpenAI implodes.

If OpenAI collapses, the ripple effects would devastate the entire tech ecosystem. Partners like CoreWeave and Oracle would be left with empty data centers and no tenants. NVIDIA could lose a large chunk of its revenue. The entire generative AI narrative that has driven tech stocks to record heights would unravel.

This raises profound questions about the resources we're pouring into AI entertainment.

Each whimsical portrait transformation is powered by technology consuming vast amounts of energy and capital with no path to sustainability. While we marvel at AI's creative abilities, the industry has failed to create viable business models that could make these technologies truly sustainable.

As we scroll through our feeds filled with AI-generated amusements, these hidden costs are worth considering.

Those delightful pet transformations represent not just technological creativity, but an industry operating on borrowed time and borrowed money. The financial structures supporting our AI entertainment aren't just inefficient — they're actively wasteful at a scale that's difficult to comprehend.

So the next time you're tempted to generate 50 variations of your dog as a superhero, remember: behind those playful pixels lies an industry burning cash at a rate that makes even the most extravagant tech bubbles of the past look restrained by comparison.

The true waste might not be in the frivolous images we create, but in the unsustainable economic foundation supporting the entire generative AI ecosystem.

Welcome to the Blacklynx Brief !

AI News

• OpenAI just launched o3 and o4-mini, its smartest reasoning models yet, now capable of using all ChatGPT tools and integrating visual thinking into problem-solving. The models push state-of-the-art performance across coding, math, and science, while the new open-source Codex CLI agent gives developers a powerful reasoning assistant directly in the terminal.

• Microsoft’s Copilot Studio now lets AI agents use apps and websites like a human — clicking, typing, and navigating interfaces without APIs. The new 'computer use' feature automates GUI tasks with built-in reasoning to adapt to interface changes, expanding how businesses can apply AI to legacy systems and everyday workflows.

• Anthropic added autonomous research and Google Workspace integration to Claude, giving it deeper access to user files and the web for smarter answers. The upgrade allows Claude to search across emails, docs, and the internet, offering personalized, cited insights — bringing it in line with the powerful “research assistant” features of its rivals.

• OpenAI also gave ChatGPT a major memory upgrade, allowing it to automatically remember details across all conversations to deliver more personalized, relevant responses. The AI will now keep track of user preferences, interests, and even dislikes without needing to be asked, though users can opt out or edit what's remembered at any time.

• Former OpenAI CTO Mira Murati’s new startup, Thinking Machines Lab, is reportedly raising $2B at a $10B valuation — one of the largest seed rounds ever. With nearly half the founding team from OpenAI, the company aims to build customizable, widely understood AI systems, though no product details have been made public yet.

• A new Microsoft study found that even top-tier AI agents struggle with software debugging tasks, solving less than half of issues in real-world tests. Despite excelling at code generation, AI still lacks the sequential decision-making skills human programmers use — highlighting a major limitation as companies race to develop coding assistants.

• Ilya Sutskever’s startup Safe Superintelligence Inc. (SSI) has raised $2 billion at a $32 billion valuation—just months after launching—with backing from top firms like Greenoaks, a16z, and reportedly Alphabet and Nvidia. The company aims to build superintelligent AI systems with safety as the top priority, despite having no public product yet.

• Twelve former OpenAI employees have filed a legal brief supporting Elon Musk’s lawsuit, arguing OpenAI’s shift toward for-profit violates its original nonprofit mission. They claim the restructuring betrays donors and staff who joined for altruistic goals, with one former researcher calling CEO Sam Altman “a person of low integrity.”

• A new Swiss-led study shows AI can detect tuberculosis through ultrasound scans more accurately than human doctors—surpassing World Health Organization benchmarks. The real-time tool runs on smartphones and could greatly improve TB diagnosis in low-resource areas, identifying early signs even trained experts often miss.

• OpenAI just launched GPT-4.1, a new API-only model family aimed at developers, with huge upgrades in coding, instruction-following, and 1 million-token context windows. It's cheaper than GPT-4o and optimized for building large-scale tools, marking a key step toward more powerful agent-based software development.

• ByteDance introduced Seaweed, a highly efficient 7B-parameter video model that rivals much larger systems like Sora and Wan 2.1—excelling in storytelling, motion realism, and lip sync. Despite its small size, it performs exceptionally well and proves that high-quality video generation doesn’t always require massive compute.

• Google released DolphinGemma, an AI model trained to understand dolphin communication—analyzing vocal patterns and even enabling real-time interaction via a custom underwater device. Built with Georgia Tech, the model may lead to breakthroughs in decoding animal intelligence and will be open-sourced for global research.

• OpenAI is reportedly developing a social media platform built around ChatGPT’s image generation tools, aiming to rival X and Meta while capturing valuable real-time data for training. While still in early stages, the platform could help OpenAI deepen user engagement and feed its models with continuous, high-quality content streams.

• Kling AI just launched KLING 2.0 Master for video and KOLORS 2.0 for images—both showing major strides in realism, editing, and prompt control. With natural motion, cinematic visuals, and flexible image styles, these updates push Kling closer to rivaling top Western models like Google’s Veo and OpenAI’s Sora.

• UC San Diego researchers tested top AI models on the game Ace Attorney, challenging them to solve courtroom cases—and most failed. While Gemini 2.5 Pro and o1 performed best, none could fully complete the case, showing that current models still struggle with long-context, interactive decision-making in complex tasks.

Quickfire News

• Safe Superintelligence (SSI) partnered with Google Cloud to use TPU chips for powering its AI research and development.

• Google CEO Sundar Pichai announced the adoption of Anthropic’s Model Context Protocol, enabling Gemini models to access external apps and data sources.

• Canva introduced Visual Suite 2.0 and a new voice-enabled AI creative partner, capable of generating editable content across formats.

• OpenAI filed a countersuit against Elon Musk, accusing him of harassment and seeking a judge’s order to block further unlawful actions.

• OpenAI also open-sourced BrowseComp, a benchmark that tests how well AI agents can search for obscure or hard-to-find info online.

• ByteDance unveiled Seed-Thinking-v1.5, a 200B-parameter reasoning model with 20B active parameters, outperforming DeepSeek R1.

• xAI launched Grok-3 via API, pricing access at $3 per million input tokens and $15 per million output tokens.

• Writer AI released AI HQ, a new enterprise platform to build, deploy, and monitor AI agents at scale.

• Llama 4 Maverick, Meta’s unmodified release version, was listed on LMArena, where it ranked below older models like Gemini 1.5 Pro and Claude 3.5 Sonnet.

• Demis Hassabis, CEO of Google DeepMind, said the company aims to merge Gemini and Veo into a single omni model with enhanced world comprehension.

• Netflix is developing a new search experience with OpenAI, enabling users to search content based on parameters like mood.

• OpenAI introduced a Verified Organization status, which will be required to access its advanced APIs and models.

• Sam Altman confirmed OpenAI plans to release a near-frontier open-source model, expanding its public model offerings.

• xAI began rolling out a memory feature for its Grok assistant, shortly after OpenAI introduced a similar tool for ChatGPT.

• NVIDIA announced its first U.S.-based AI manufacturing initiative, partnering with TSMC, Foxconn, and others to produce chips and supercomputers in Arizona and Texas.

• OpenAI is expected to release two new models this week—o3 and o4-mini—which will reportedly be capable of generating scientific ideas and automating advanced research.

• Amazon CEO Andy Jassy stated in his shareholder letter that generative AI will “reinvent virtually every customer experience” in the future.

• Meta plans to train AI on public content from EU users, while offering an opt-out form and emphasizing the need to reflect European culture in its systems.

• Hugging Face acquired Pollen Robotics and released Reachy 2, an open-source humanoid robot priced at $70,000 for research and AI embodiment use cases.

• LM Arena debuted a Search Arena Leaderboard, ranking LLMs by search capabilities, with Gemini-2.5-Pro and Perplexity’s Sonar leading.

• NATO awarded Palantir a contract to deploy its Maven Smart System, enhancing battlefield operations with AI within 30 days.

• OpenAI updated its Preparedness Framework, stating it may revise safety requirements if competitors begin releasing high-risk AI systems without equivalent safeguards.

• ChatGPT now features a Library tab, giving free and paid users access to all their AI-generated images in one place.

• xAI launched Grok Studio, a collaborative workspace similar to ChatGPT’s Canvas, allowing users to work with Grok on documents, code, games, and reports.

• Cohere released Embed 4, a multimodal embedding model with a 128K context window, 100+ language support, and potential 83% storage savings.

• Google released Veo 2, its top-tier video generation model, to Gemini Advanced, AI Studio, and Whisk users.

• Nvidia disclosed a projected $5.5B loss due to U.S. export license restrictions affecting shipments of H20 AI chips to China.

• Microsoft added computer use features to Copilot Studio, allowing users to build agents that interact with desktop and web apps using UI-level actions.

• OpenAI is in acquisition talks to buy Windsurf (formerly Codeium), a coding platform, in a potential deal worth up to $3B.

• Microsoft researchers introduced BitNet b1.58 2B4T, a 1-bit AI model that performs on par with larger models while being efficient on CPUs.

• Tencent launched FireEdit, an AI image editing tool that uses region-aware vision language models for precise, instruction-driven edits.

• Anthropic is preparing to debut a new voice mode for Claude, featuring three AI voices named Airy, Mellow, and Buttery, expected this month.

• Metr released an evaluation of OpenAI’s 3o and 4o-mini models, noting the safety testing was conducted on an accelerated timeline.

• Tyler Cowen stated that o3 may qualify as AGI, suggesting April 16 could be viewed as the date of artificial general intelligence arrival.

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 !