Enterprise IT leaders are increasingly uncomfortable realizing that Generative AI (Genai) technology is still in progress, and buying it is like spending billions of dollars to take an Alpha test – not even a beta test, but a Early Alpha, coders can hardly keep up with error reports.
For those who remember the first three seasons Saturday Night Live,Genai is the final algorithm that is not ready for the algorithm.
One of the latest evidence comes from Openai, which had to shockingly withdraw the recent version of Chatgpt (GPT-4O), when it (among other things) provided a very inaccurate translation.
Lost in Translation
Why? exist CTO’s words Who found this question: “Chatgpt doesn’t actually translate the documentation. It guesses the sounds I want to hear, blending it with past conversations to make it feel reasonable. It not only predicts words. It predicts what I expect. It’s absolutely terrible, like I really believe.”
Openai says Chatgpt is so good.
“We rolled back the GPT-4O update in Chatgpt last week, so people are now using earlier versions of more balanced behavior. The update we removed was too flattering or pleasant – often described as Sycophantic, Openai explainedAdded that in the “GPT-4O update, we adjusted to improve the default personality of the model to make it feel more intuitive and more effective in various tasks. We focused too much on short-term feedback and did not fully take into account the user’s interaction with Chantgpt, and the results evolved over time. As a result, GPPT-4O tends to respond, whereas responding, which are too supportive but not responding in time.
“…Each of these ideal qualities (such as trying to be useful or supportive) has unexpected side effects. 500 million people use Chatgpt every week, and in every culture and environment, a default value cannot capture all preferences.”
Openai is intentionally blunt. The problem is not that the app is too polite and manageable. This is not a question of imitating the lady’s behavior.
I am no If you ask me to translate the document, I will tell you what sound I want to hear. This is similar to Excel taking away your financial data and making net income bigger because it thinks it will make you happy.
Just like the way IT decision makers expect Excel to calculate numbers accurately, no matter how it affects our sentiment, they expect translations of Chinese documents won’t constitute something.
Openai cannot say “trying useful or supportive ideal traits can have unexpected side effects.” Let’s be clear: giving others the wrong answer will have accurate expected results – the wrong decision.
Yale University: LLM requires data marked as error
Alas, Openai’s happiness efforts aren’t the only weird Genai news lately. Yale University researchers explore a fascinating theory: If the LLM trains only the information marked as correct – whether the data is actually correct or not is not a substance – it has no chance to identify defective or highly unreliable data because it does not know what it looks like.
In short, if you have never been trained for data that marks it as False, how could it be able to recognize it? (this Complete study at Yale University Here. )
Even the U.S. government found that Genai’s claim was too far away. It’s a big statement when the Fed lies are too far away.
FTC: Genai supplier makes false, misleading claims
The Federal Trade Commission (FTC) found that a large language model (LLM) supplier, Workado, Flawed claims about the accuracy of their LLM detection products. It expects suppliers to “maintain competent and reliable evidence that these products are as accurate as claimed.”
Chris Mufarrige, director of the FTC Consumer Protection Agency, said customers “trust Workado’s AI content detector to help them decipher whether AI is behind writing, but the product is no better than a coin throwing away.” “The misleading nature of AI creates a more difficult misunderstanding of legal providers of AI-related products that are more difficult to attract consumers.
“…The order resolves the allegation that Workado has promoted its AI content detector to “98%” because it can detect whether the text is written by AI or humans. But independent tests show that the accuracy of universal content is only 53%,” FTC’s Administrative Complaints.
“The FTC claims Workado violates the FTC Act because 98% of claims are false, misleading or unfulfilled.”
Here is a key lesson. Genai suppliers make significant claims for their products without meaningful documentation. Do you think Genai has created something? Imagine what the supplier’s marketing department will bring.
