Today's "experts"--influencers, journalists--repeat press releases as if they were commandments. They seek FOMO, selling their course/service/advertising, and racing to gain a few followers.
Multinationals put on the show: they launch, announce wonders, and increasingly fail to deliver. The problem is not one company; it is the launch industry.
On launch night, I got up early to test GPT-5 thoroughly.
I was surprised by the notorious instability, confused reasoning (at unimaginable levels), and above all the degradation in GPTs, projects, agents, and automations that previously worked.
Inexplicably, for 24 hours social media remained positive. "Copying and pasting" the same messages. Then the tide of opinions began to reverse.
What changed? Tweets from Sam Altman acknowledging problems and, once again, the echo of press releases.
It is not just OpenAI. They sought to unify and simplify and launched with significant errors. Google overuses grandiose announcements with Gemini: headlines today, explanations tomorrow. The problem is not one company: it is the launch industry.
What is the underlying problem?
This logic of the "big announcement" is alien to modern innovation practices. In lean and agile, you learn with business stakeholders: you experiment and co-create, you deploy in small batches to production, you test with real samples, you measure and refine. Here, instead, the show is prioritized and the risk is transferred to the user.
That old--very commercial--logic (the need to "show" market upturn, fight for headlines) is creating serious risks for companies and users.
OpenAI promises consistency and simplicity, but the execution leaves performance problems and uncertainty. Google mocks up promises that it then fails to deliver or that do not work as announced. Result: the positioning and value proposition of both erode, and trust in the application of these tools falls.
Competition for headlines. Some announce too early; others launch unstable. In both cases, the user pays--with time, errors, and frustration--while organizations must contain the impact.
When the model changes, your operation pays. You build internal GPTs, assistants, and agents for specific tasks--classifying tickets, responding to customers, generating reports, triggering orders--and overnight, ChatGPT unilaterally changes how they "think." Result:
Formats that no longer meet what your system expects (empty fields, different names, new structures).
Rules that contradict themselves: the same case, two different decisions.
Truncated sequences in agents: they skip steps or reverse them.
Automations with impact on users or physical operations (shipments, reservations, billing) that fail in cascade.
That chaos is not only paid for in hours: but in rework, escalations, apologies, and reputation.
Yes, there are technical barriers to reduce these risks, but they are not infallible: models can change or be retired, drift in their behavior, vary in response times and costs, or fail unexpectedly, impacting customers, operations, and brand. The only responsible stance is to design assuming that uncertainty, build managing risks, and decide with continuous evidence, not with headlines.
Our interaction with LLMs is not only functional; it is relational and emotional. Over time, we build trust with an assistant: its tone, its way of reading, interpreting, and its intuitions. When that is maintained, flow appears and work moves smoothly.
If that "personality" changes overnight and you cannot protect or choose it, value is destroyed:
Adoption drops: people stop using it because "it no longer understands me."
Errors from expectations increase: the team believes the assistant will do X and now it does Y.
Trust breaks: every response requires double-checking; you return to manual work.
What the industry could/should offer (beyond uptime):
Behavioral stability as a commitment: not changing tone/style/format without notice.
Notices and windows before modifying behavioral traits, with the option to temporarily maintain the previous model to transition without breaking workflows.
Sustainable value and trust = operational robustness (nothing breaks when the provider changes) + continuity of personality and interaction (the assistant "remains itself").
Do not outsource judgment. Business intelligence/strategy is not outsourced; neither is AI, nor continuous experimentation.
Learn in batches with the user at the center. Small pilots, clear metrics, and decisions based on evidence, not headlines.
Design for change. The question is not whether the model will change, but how you will absorb that change without harming your users and business stakeholders.
Snake oil sellers: they are exposed when their narrative is not based on experimentation and real knowledge. Do not follow applause; demand judgment. And build your own.
Press and "experts": their asset is credibility. Copying and pasting destroys that capital.
Big tech companies: their value depends not only on what they promise, but on how they execute change without harming the user (or their bond with the assistant).
User companies: the advantage lies in integrating, managing risk/return, and learning better than the rest.
With fresh data: My tests in handling long texts (editing chapters and extensive documents) changed completely compared to yesterday: the results are impressive in performance and speed.
