hallucination

"Instead of making the AI smarter, it's about building a cage around the AI. This means shifting from Human-in-the-Loop to Program-in-the-Loop."

"ChatGPT lies because it is fundamentally built in a way that 'it is more beneficial to lie.'"

"I replaced LLM knowledge storage with a database and it works. Long live LLMs (without the Hallucinations)."

"For those working with Large Language Models (LLMs) in production, what kinds of questions should I expect? Specifically: System design: Do they ask you to design things like RAG pipelines or LLM-based applications?"

"I feel like people say AI is amazing, but when it actually matters, they: cross-check, Google after, ask another AI"

"Generative AI creates "plausible-looking text," so even if it is wrong, it can output natural lies. In a word: "Plausibility ≠ Correctness.""

"If you have LLM write Prolog and leave the execution to Prolog... I created prolog-reasoner to enable SWI-Prolog to be used as an MCP server."

"As a result, cases where 'the meaning is correct, but it is slightly misaligned with our company's past answer policy' occurred with conventional AI, and the issue was that man-hours for微修正 the outputted answers remained."

"According to this paper, increasing thinking tokens (inference tokens) improves accuracy, but beyond a certain point, the improvement plateaus."

"From my experience, this new adaptive reasoning system seems to default to a low-effort mode for most queries to save compute."

"So, cyber security of tomorrow will not be like proof of work in the sense of 'more GPU wins'; instead, better models, and faster access to such models, will win."

"The intelligence of AI resides in a world of 'snapshots' where everything is fixed the moment training ends. They structurally lack the concept of 'now' and are always living within 'past records'."

"Basically, it gives the network a mathematically rigorous 'I don't know' button for free."

"Bridgebench is accusing Anthropic of last week Claude Opus 4.6 ranked #2 on the Hallucination benchmark with an accuracy of 83.3%. Today Claude Opus 4.6 was retested and it fell to #10 on the leaderboard with an accuracy of only 68.3%."

"However, in reality, even amateurs like myself can deploy a complex, cross-platform application that was a long-held dream, and experience that intense thrill."

"Instead of having AI do everything, we made the decision to strip away features for practicality, focusing on '80% accurate analysis across all thousands of records' rather than '100% accurate analysis on just 10 records'."

"Large language models (LLMs) can struggle to memorize factual knowledge in their parameters, often leading to hallucinations and poor performance on knowledge-intensive tasks."

"As a result of thoroughly exploring Hallucination countermeasures, the next step is perhaps to prohibit the AI from trying to forcibly use the same words."

"AI is trained by default to answer 'helpfully and comprehensively.' That is exactly why we need to explicitly apply the brakes and tell it to 'make it short.'"

"Social pleasantries (greetings, apologies, closing words) are completely unnecessary. Always keep answers concise. State the conclusion first."

"However, in multiple practical business scenarios, I repeatedly encountered situations where Gemini confidently provided misinformation about Google's own services and, even when pointed out, did not retract it."

"I really think the golden age of consumer and prosumer access to LLMs is done."

"Because AI is trained by default to answer 'helpfully and comprehensively,' it is necessary to explicitly apply the brakes and tell it to 'be brief.'"

"LLMの演算は高次元ベクトル空間における非線形変換の反復です。自己回帰的な生成方式では直前の出力が次の入力となるため、微小な不確定要素が累積し、終着点の予測は不可能になります。"

"claims of 'thousands' of severe zero-days rely on just 198 manual reviews"

"Google DeepMind's research team has for the first time systematically classified how malicious web content can 'weaponize' AI agents."

"It acts as a 'veteran genius super-employee': an existence that works as a central command tower for specialized AI agents (legal, accounting, etc.) by comprehensively understanding the company's vast amount of internal data."

"LLMs have a knowledge cutoff and no native access to live information. Enable search. Give it context."

"I sent it an image that featured a comb and asked it if it could find that exact same comb for cheaper anywhere else... and gemini started telling me about the history of cactuses"

"M2.7 is Intelligence Index 50, equivalent to GLM-5 (Reasoning), but the total cost to run the full index is $176, less than 1/3 the cost of GLM-5."