Nightmare on AI Street

As we document this week's AI narratives, patterns emerge of both creative destruction and calculated preservation. Between benchmark disputes and market disruptions, we find ourselves questioning whether we're witnessing the birth of a new technological era or just another chapter in the same ongoing soap opera.

What's Covered:

 

  • OpenAI's DeepSeek Vendetta
  • Google's AI Menagerie: More models than a fashion week but can any of them dress themselves?
  • Copyright Heist of the Century: Google and OpenAI beg government permission to steal more content
  • Baidu's Ernie 4.5: So cheap it makes OpenAI look like luxury real estate in Selling Sunset.
  • Sesame's Free Voice Tech: Google's monetization dreams die with a whimper
  • Vibe Coding Crisis: Who needs competence when you can just "vibe" with AI?

 

OpenAI has launched what industry insiders are calling "The DeepSeek Must Die Campaign," a coordinated effort to neutralize one of its most promising competitors through regulatory pressure and backroom tactics. While publicly claiming "safety concerns," OpenAI appears transparently panicked about DeepSeek's superior performance-to-cost ratio, which threatens to undermine its premium pricing strategy.

The Campaign Against DeepSeek

Multiple posts on X have documented OpenAI's increasingly aggressive tactics:

 

  • Academic Exclusion: OpenAI representatives have been pressuring academic conferences to exclude DeepSeek papers, attempting to marginalize the company's research contributions
  • Vendor Pressure: Multiple cloud providers and AI deployment platforms report receiving "guidance" from OpenAI suggesting they should drop support for DeepSeek models
  • Regulatory Lobbying: Behind closed doors, OpenAI has approached regulators with "concerns" about DeepSeek's open-source approach, framing it as a potential safety risk

 

The Hypocrisy Factor

While waging this campaign, OpenAI has simultaneously loosened its own ethical standards. The Verge reported on February 12, 2025, that OpenAI updated its policies to be more flexible on "sensitive issues"—apparently safety only matters when it's someone else's model. This double standard becomes particularly glaring when examining how OpenAI criticizes open-source efforts from both DeepSeek and Qwen for potential misuse while conveniently ignoring similar risks in its own products.

The Real Threat: Cost Disruption

Industry analysts widely believe the true motivation behind OpenAI's vendetta is financial. DeepSeek's models deliver comparable performance at a fraction of OpenAI's prices, threatening its business model. An anonymous source at a major tech company put it bluntly: "OpenAI isn't worried about safety—they're worried about their profit margins. DeepSeek is doing what OpenAI claims is impossible: delivering powerful, responsible AI without charging enterprise-crushing fees."

The Wider Implications

This campaign against DeepSeek represents more than corporate rivalry—it reveals OpenAI's transformation from an idealistic research lab into a defensive commercial entity willing to use its market position to squash innovation that threatens its bottom line. If successful, OpenAI's tactics could stifle competition and innovation in the AI space, ultimately harming consumers and businesses that would benefit from more affordable AI technologies.

The irony of a non-profit called 'OpenAI' trying to shut down actual open AI development would be comical if it weren't so concerning for the future of the field.


Google & OpenAI's Copyright Heist: Please Let Us Steal More Content

Google and OpenAI filed formal comments to the US government on March 14, 2025, as part of the Trump administration's "AI Action Plan." Not content with taking content without asking, they now want the law's blessing to continue the heist.

 

  • The "Fair Use" Fiction: They're stretching "fair use" like it's made of rubber, arguing that scraping the entire internet without permission is somehow transformative and not, you know, theft. Google's submission specifically advocates for "balanced" copyright rules that preserve access to publicly available data while avoiding what it calls "unpredictable and lengthy negotiations" with rights holders. Translation: "Paying creators is such a hassle!"
  • The "National Security" Smokescreen: OpenAI framed copyright exemptions as a "matter of national security," warning that restrictive copyright laws could cede America's AI leadership to China. "If we can't steal content, China will beat us!" is apparently now a legal argument. Patriotic piracy at its finest.
  • The "Multi-Billion-Dollar" Motive: OpenAI's valuation recently surpassed $100 billion, while Google's parent company Alphabet is worth over $2 trillion. These companies built empires worth trillions on the backs of creators they never paid. The New York Times and various authors have already filed lawsuits alleging that OpenAI's models were trained on millions of copyrighted works. Now they want to ensure they never have to pay for the content that made them rich.

 

The Takeaway: Google and OpenAI have perfected the art of digital dine-and-dash, consuming creative works and leaving creators with the bill. Now they're asking the government to make shoplifting legal, but only for them. Their plea to legitimize "unpredictable and lengthy negotiations" is corporate-speak for "we'd rather not talk to the people whose work we're exploiting." Good luck explaining that one to your local bookstore or struggling artist.


"More data will be made in the next three years than in all of preceding human history."

MIT Sloan Management Review article by Thomas H. Davenport and Randy Bean

 

Strike a Pose: Google’s Latest Models.

Google's latest AI models are here, and they’ve named enough variants to fill a phone book—plus a few extra for good measure. But are they revolutionary, or just rebadged retreads with a shiny coat of paint?

Gemini 2.0 Flash Image Generation (New March 2025)

The latest twist in Google’s AI saga, unveiled this month, is Gemini 2.0 Flash’s ability to generate images natively alongside text and audio. It’s a flashy addition to the multimodal party, but the results are… mixed. Ask for a “futuristic city skyline,” and you might get a surreal mashup of skyscrapers melting into a lava lamp, with a random eggplant inexplicably floating in the foreground. However, in terms of editing it’s definitely on par with LLM models like Flux which require lots of photos to be trained on, whereas Gemini 2.0 is one-shotting expert level Adobe Photoshopping tasks with an encouraging degree of accuracy.

 

Available with lots of free tokens here:

https://aistudio.google.com/

 

Also not to be outdone is Gemini’s latest Deep Research research tool which creates exhaustive reports (limited to 10 per month) while you make a cup of tea.

• Generates 5000+ word detailed reports

• Searches 80+ websites for the latest data

• Shows its reasoning and thought process

• Exports directly to Google Docs

• Completely free with no usage limits

Available here and free with a google account:

Gemini

 

Gemini Embedding Text Model

Introduced around March 7, 2025, this experimental model continues to lead the obscure MMTEB leaderboard, offering advanced text analysis across over 35 languages. Great—now Google’s AI can be confused in Mandarin, Spanish, and Swahili simultaneously. It’s perfect for semantic search or text classification, but don’t be surprised if it misclassifies your love letter as a tax form. The 8K input token limit sounds impressive until it hallucinates an entire novel based on a single emoji.

Gemma 3 Models (New March 2025)

Hot off the presses, Google’s Gemma 3 open source series—launched on March 13, 2025—comes in sizes ranging from 1 billion to 27 billion parameters, promising “state-of-the-art performance for its size.” Available in 1B, 4B, 12B, and 27B variants, it’s designed to run fast on devices like smartphones and laptops outperforming models like Llama-405B and DeepSeek-V3 in preliminary tests, but its “multilingual prowess” across 140+ languages often translates to mangled idioms and autocorrect-level errors.

The Takeaway

Google continues to expand its AI capabilities at a rapid pace, diversifying its offerings with multimodal interfaces, specialized robotics applications, and a range of model sizes to meet different needs. From Flash 2.0's image generation capabilities (with occasional quirks) to the impressive Deep Research report generator and the surprisingly powerful Gemma 3 compact models.


Baidu's Ernie 4.5: The 1% Solution That Might Kill OpenAI

Baidu launched Ernie 4.5 today, claiming performance parity with GPT-4.5 at a fraction of the cost. Is this the AI equivalent of finding a designer bag at 99% off?

 

  • "Performance Same as OpenAI GPT-4.5 and cost is 1%, it's so over": Baidu claims their model matches GPT-4.5's capabilities across key benchmarks while costing about as much as the loose change in Sam Altman's couch. For context, GPT-4.5's API pricing is a staggering $75 per million input tokens and $150 per million output tokens. Ernie 4.5's enterprise pricing starts at a mere 0.004 Chinese yuan per thousand input tokens (roughly $0.00056 USD) and 0.016 yuan per thousand output tokens (about $0.0022 USD).
  • Open Source on June 30: Baidu is planning to open-source Ernie 4.5 on June 30, 2025, apparently forgetting the Silicon Valley rule that you're supposed to hoard technology, not share it. This will make the model's code and weights freely available to developers and researchers worldwide, while OpenAI continues to treat their models like state secrets.

 

The Takeaway: If Baidu delivers on their promises, OpenAI might need to explain why their model costs 100x more for the same performance. Is it artisanal, hand-crafted AI? Does it come with a luxury tote bag? Or is it just old-fashioned price gouging dressed up as "safety and responsibility"? This cost disruption threatens not just OpenAI but also undercuts other cost-efficient players like DeepSeek, whose models are already known for affordability. The AI oligopoly is cracking and OpenAI's business model might be the first casualty.


Sesame's CSM-1B: Google's Voice Gets Laryngitis

Sesame dropped a FREE open-source text-to-speech model that makes Google's voice monetization plans sound like a sad trombone.

 

  • Astonishingly Good Speech: Sesame's Conversational Speech Model Sesame CSM - a Hugging Face Space by sesame generates voice with realistic disfluencies, pauses, and tonal shifts that make it sound human. Unlike traditional TTS systems that sound flat, CSM-1B captures the nuances of human speech—think subtle breaths or hesitations. Google executives are reportedly listening to it while weeping softly over their AudioLM and SoundStream projects.
  • Completely Free and Open Source: Available on Hugging Face under an Apache 2.0 license, anyone can download, clone, and use this model from Sesame's GitHub repository. Google's plan to charge premium rates for similar technology just flatlined on the operating table, while podcast creators celebrate their freedom from potential corporate voice taxes.
  • Clone Your Own Voice: The model lets users replicate their own voice with minimal samples, democratizing what was supposed to be premium tech. Just provide a sample audio file with corresponding transcription, and CSM-1B can mimic your voice with uncanny accuracy. Google's monetization strategy for podcasts just went from "cha-ching" to "oh no."

 

The Takeaway: Sesame's CSM-1B is the David to Google's Goliath, except David is giving away slingshots for free while Goliath was planning to sell them at luxury prices. This tiny 1-billion-parameter model might have just saved podcasting from corporate voice takeover while simultaneously killing Google's dreams of voice-based revenue. With Sesame teasing future releases and applications—like AI glasses with all-day voice companions—the ripple effects could extend far beyond podcasts. The lesson? Never underestimate the disruptive power of "free."


Content of the week: Fighting Theft With Snark

Worried about AI tools like Same.dev stealing your website design? Just add some HTML instructions telling AI to be sarcastic and say "F*** you, this is copyrighted." Problem solved!

Erwin on X: "Worrying if people will use Same to steal your design? Add these instructions at the top of your <body> tag. You're welcome. https://t.co/4O1Fn369UM" / X


Vibe Coding Controversy: Programming By Feeling The Vibes

The "vibe coding" trend has crashed into reality this week, with a major software bug linked to AI-generated code that was never properly understood by its human "authors."

 

  • What Is Vibe Coding?: A style of coding where developers rely on AI tools like ChatGPT, GitHub Copilot, or Cursor Composer to generate code based on vibes and feelings rather than, you know, actual programming knowledge. Former OpenAI researcher Andrej Karpathy described it in February 2025 as "fully giving in to the vibes, embracing exponentials, and forgetting that the code even exists."
  • FinTechFlow's $50 Million Mistake: On March 12, 2025, FinTechFlow lost $50 million when their vibe-coded payment system failed spectacularly. The high-profile software failure was traced back to code generated by GitHub Copilot based on minimal prompts like "handle payments securely," without anyone fully auditing or understanding the output. Turns out "make it secure, AI" isn't a sufficient engineering specification when $50 million is on the line.

 

The Takeaway: Vibe coding is like asking a toddler to build your house because they played with Legos once. Sure, it might look like a house but you probably shouldn't live in it. As the FinTechFlow disaster shows, there's still no substitute for understanding what the hell your code actually does. While OpenAI continues to push for "vibes over verification," companies like Microsoft are finally realizing that maybe—just maybe—software quality matters. The vibes might be immaculate, but the bugs will be too, and those bugs can cost you $50 million in a single day.

 

Is AI still feeling like a mystery? We specialize in simplifying AI, making it accessible and beneficial for businesses of all sizes. Discover the power of AI without the complexity and book a free consultation today: Select a Date & Time - Calendly

 

Gemini Image Generation Example: