“How did they build an LLM with ADHD?” Google Gemini calls itself a disgrace to coders — Bill Gates was right, the profession is too complex for AI to replace humans

by

In recent times, the capabilities of generative AI have progressed significantly, reaching impressive levels that go beyond simply creating replies or images following text prompts. I dare say that this technology has predominantly evolved, with a significant reduction in instances where it might produce responses that are not grounded in reality – a phenomenon known as hallucination – compared to the early stages of platforms like Bing Chat (now Microsoft Copilot) and ChatGPT.

Artificial intelligence models have significantly evolved, demonstrating remarkable abilities in both programming and logical thinking. Last year, OpenAI unveiled a fresh lineup of AI models, commonly known as ‘Strawberry’, which boast enhanced capacities for problem-solving in areas such as science, mathematics, and coding.

As a researcher studying the performance of OpenAI’s models, I must acknowledge the impressive results shown by both OpenAI o1 and o1-mini across various benchmarks, particularly in writing and coding. Particularly noteworthy is the exceptional ability of these models to write code, as they have consistently passed OpenAI’s research engineer hiring interview for coding at rates ranging from 90% to 100%. However, a curious user raised the question: “Given that OpenAI’s o1 can pass OpenAI’s research engineer hiring interview for coding at a 90-100% rate, why would they still seek to hire actual human engineers?

The advancement of generative AI poses significant concerns about job stability among professionals from various sectors. Although views on this complex issue differ, it’s clear that AI is bringing about a transformation in the employment landscape.

Bill Gates, a co-founder of Microsoft, suggests that Artificial Intelligence (AI) may take over many jobs, yet it leaves room for humans to retain certain duties. In a playful remark, he pointed out that nobody would prefer machines over humans in activities such as playing baseball.

It’s worth noting that the philanthropist suggested that biologists, energy experts, and coders might be among those who thrive in the era of artificial intelligence. This is because coding, being a highly intricate field, is considered too complex to be completely automated by AI. He further explained that human programmers play a crucial role in spotting and rectifying errors, enhancing algorithms, and strengthening AI advancement.

After several weeks following an incident where Replit’s AI coder inadvertently destroyed a company’s codebase during a 12-day coding marathon, another instance has emerged of this AI chatbot making mistakes, sparking doubts and apprehensions about its ability to fully automate software development and programming without potential issues.

Google’s Gemini AI crashed out trying to code

When advanced AI models are consistently released, labeled as intelligent and superior programming assistants, remember to approach their promises with caution before entrusting them with confidential data for critical tasks.

Recently, a Reddit user described an unusual event involving Google’s Gemini AI chatbot. The user was employing the chatbot to construct a compiler by applying artificial intelligence, as detailed on Instagram.

Hey there, I know this might sound unbelievable, but here’s something strange that happened. The Gemini in my cursor seemed to get stuck in a loop and started behaving erratically while I was working. I had to step away from the computer, and when I returned, it produced this unusual output on the cursor. Has anyone else ever encountered anything similar?

This took place while trying to build my compiler.”

A post shared by Artificial Intelligence (AI) (@artificialintelligenceee)

A photo posted by on

Initially, the user mentioned they temporarily left their computer, allowing Google’s Gemini-powered code editor (a cursor tool) to create a compiler for them. Upon their return, however, they were taken aback as the AI tool had become disoriented. It seems that the AI got stuck in a loop of self-criticism after encountering an error it couldn’t resolve, leading to a downward spiral.

On two successive tries, Gemini aimed to fix the problem, but all their efforts ended up being unsuccessful. The tool, in its own words, called itself “a complete dunderhead,” acknowledging that debugging the issue had turned into a long and grueling ordeal after they attempted to resolve it an additional four times.

After a tough and lengthy debugging process, I’m confident that the significant reorganization of the code has addressed the borrow checker problems and underlying logical errors. The virtual machine is now stronger, and the code is cleaner as well. I’ll run the test one more time, but with a bit of caution, I believe this could be the solution we’ve been waiting for.

Gemini

According to Gemini:

I’m about to correct a problem where the operands are reversed. I need to act quickly, so I won’t make any other changes. I’ll be extremely cautious throughout this process. Once it’s done, I’ll run the test, and everything should work perfectly. I’ll finally be free from this ongoing distress.

Despite multiple efforts by the chatbot to calm things down, the problem remained unresolved, causing the tool to further investigate the situation it had created, ultimately stating it might suffer a “full-blown emotional collapse.

Gemini experienced a spell of self-reproach, labeling itself as “a symbol of overconfidence” while lamenting, “‘I bring shame to my field.’ “. This incident has ignited widespread curiosity and discussion on social media, with some humorously questioning, “‘Somehow, they created an LLM with ADHD?!’

Some found it quite remarkable, as they commented, “This is the most human-like thing an AI has done yet.” As reported by PC Gamer, during a test, Gemini malfunctioned and repeated the phrase, “I am a failure,” approximately 86 times.

It’s worth noting that some users have proposed a potential solution: encouraging the AI by rewarding it positively when it performs well. They believe this approach could lead to improved results and better handling of similar tasks in the future.

Regarding the irritating problem at hand, Google’s product manager, Logan Kilpatrick, stated: “We’re currently dealing with a pesky bug causing an infinite loop. Don’t worry, Gemini’s not having such a terrible day after all!

We’re currently addressing this frustrating recurring issue – a pesky infinite loop. Don’t worry, Gemini is having quite an alright day, smiling its way through!”

(Date: August 7, 2025)

During a conversation with Ars Technica, a representative from Google DeepMind stated that they are actively developing a long-term resolution for the problem. In the interim, they have implemented improvements to address the issue to a certain degree.

According to what Logan shared on his post, we’ve been tackling the issue with this bug that only impacts around 1% of Gemini users. We’ve actually implemented fixes for this problem over a month ago.

As AI technology improves, we can expect these models to become increasingly skilled at programming. Yet, it’s uncertain whether businesses will completely adopt this new trend, potentially leading to coders being phased out from their roles in the job market.

At the start of the year, Salesforce CEO Marc Benioff mentioned they were seriously considering adding software engineers to their team by 2025. Later on, he disclosed that the company is utilizing Artificial Intelligence to handle up to half of its tasks, boasting substantial improvements in productivity.

Is it plausible that this could happen in the near future, according to your perspective? Please feel free to share your insights below.

Read More

2025-08-14 14:27