The ELECTE blog
Back to the front page All articles
Newsletter
Opportunities for AI startups in 2025 *UPDATED*
"While everyone is scrambling to implement GPT-5, some people are still making money selling buttons." The real AI opportunity in 2025 is not reinventing the wheel-it's solving real problems without burning up budgets. Undervalued niches: personalization that doesn't make customers feel like they're in Black Mirror, healthcare assistants that distinguish a cold from the emergency room, analytics for SMBs that hate Excel. Success? Not of those with the most powerful AI, but of those who make it accessible, useful and sustainable.NewsletterThe New Gold Rush: History, Comparisons and Future Prospects
Klondike 1896: 100,000 people set out for the Yukon, few found gold-the winners were those who sold shovels. AI is a new gold rush, but with crucial differences: demand outstripping supply (not the other way around as in the dot-com bubble), immediate economic value, financially sound companies. We are at the 1995-98 equivalent of the Internet. The historical lesson? Intermediate technical skills are short-lived, domain knowledge retains value. Better to sell shovels or pan for gold?NewsletterThe Illusion of Progress: Simulating General Artificial Intelligence Without Achieving It
We are not building AGI-we are building an increasingly convincing illusion. In 2025, general intelligence will emerge not from a single system, but from a mosaic of coordinated specialized AIs: LLMs, image generators, AlphaFold. Quantum computing promises to exceed the computational plateau (-99% consumption according to IBM), while Microsoft and Google compete with radically different approaches. The provocation? If human consciousness is itself an emerging illusion, perhaps AGI "by proxy" is more like us than we think.NewsletterThe Strawberry Problem
"How many 'r's' in strawberry?" - GPT-4o answers "two," a six-year-old knows it is three. The problem is tokenization: the model sees [str][aw][berry], not letters. OpenAI didn't solve it with o1-it got around it by teaching the model to "think before you speak." Result: 83% vs. 13% in Math Olympiad, but 30 seconds instead of 3 and triple the cost. Language models are extraordinary probabilistic tools-but you still need a human to count.NewsletterThe Great Deception: Why AI Understands Emotions Better Than It Admits
82% AI vs. 56% human accuracy in emotional intelligence tests-the Geneva and Bern study demolished our last reassuring myth. ChatGPT-4 not only outperforms humans in existing tests-it creates new ones indistinguishable from those of professional psychologists. Microexpressions, speech analysis, contextual understanding-AI reads emotions we ourselves do not recognize. The question is no longer "can it understand emotions?" but "how do we harness this superior understanding while keeping human values at the core?"NewsletterEvolution of LLMs: a brief overview of the market
Less than 2 percentage points separate the top LLMs on key benchmarks-the technology war ended in a tie. The real 2025 battle is played out on ecosystems, deployment, and cost-DeepSeek proved it can compete with $5.6M vs $78-191M of GPT-4. ChatGPT dominates brand (76% awareness) despite Claude winning 65% of technical benchmarks. For companies, the winning strategy is not to choose "the best model" but to orchestrate complementary models for different use cases.Page 3 of 3