Why I don’t think AGI is right around the corner
Things take longer to happen than you think they will, and then they happen faster than you thought they could
OpenAI Hits the Panic Button
The company behind ChatGPT is preaching mission over money as Meta poaches its staff. But a forced vacation and a major departure at a friendly startup tell a different story.
AI threatens world with 'Mad Max' jobs market warns top economist
MIT economist David Autor said on the "Possible" podcast that AI may lead to a world where jobs remain but skills are devalued.
'I'm being paid to fix issues caused by AI'
Businesses that rush to use AI to write content or computer code, often have to pay humans to fix it.
What I learned building an AI coding agent for a year
We thought we'd be the best within months. Here's what went wrong — and why I'm more excited than ever.
The Rise of Whatever | Hacker News
There is at least three major art markets: 1) pretty pictures to fill in a void (empty walls, dress up an article...), 2) prestige purchases for those trying to fill that void in their imposter syndrome, and 3) fellow artists who are really philosophers working beyond language. The whole reason art is evaluated with vague notions like taste, context, history and so on is because the work of artists left their audience's understanding several generations ago, but they still need to make a living, so these proxies are used so the general public does not feel left out. Serious art is leading edge philosophy operating in a medium beyond language, and for what it's worth AI will never be there, just like the majority of people.
There Are No New Ideas in AI… Only New Datasets
LLMs were invented in four major developments... all of which were datasets
How Long Contexts Fail
Taking care of your context is the key to building successful agents. Just because there’s a 1 million token context window doesn’t mean you should fill it.
longer contexts do not generate better responses. Overloading your context can cause your agents and applications to fail in suprising ways
The New Skill in AI is Not Prompting, It's Context Engineering
Context Engineering is the new skill in AI. It is about providing the right information and tools, in the right format, at the right time.
NOVITÀ: Claude di nuovo al top con i NUOVI artefatti
🙏 Supporta il canale abbonandoti qui: https://www.youtube.com/channel/UCrebGs3b-Z7JLKQM2YOpUKA/joinAnthropic colpisce ancora con un incredibile aggiornament...
AI systems are unlikely to make the scientific discoveries some leading labs are hoping for, Hugging Face's top scientist says
Hugging Face's co-founder, Thomas Wolf, is pouring cold water on the hopes that current AI systems could revolutionize scientific progress.
Novel Universal Bypass for All Major LLMs
HiddenLayer’s latest research uncovers a universal prompt injection bypass impacting GPT-4, Claude, Gemini, and more, exposing major LLM security gaps.
One Prompt Can Bypass Every Major LLM’s Safeguards
Researchers have discovered a universal prompt injection technique that bypasses safety in all major LLMs, revealing critical flaws in current AI alignment methods.
Watching o3 guess a photo’s location is surreal, dystopian and wildly entertaining
Watching OpenAI’s new o3 model guess where a photo was taken is one of those moments where decades of science fiction suddenly come to life. It’s a cross between the …
LLMs can't stop making up software dependencies and sabotaging everything • The Register
: Hallucinated package names fuel 'slopsquatting'
All that's required is to create a malicious software package under a hallucinated package name and then upload the bad package to a package registry or index like PyPI or npm for distribution
Coding agents have crossed a chasm
Coding agents have crossed a chasm Somewhere in the last few months, something fundamental shifted for me with autonomous AI coding agents. They’ve gone from a “hey this is pretty neat” curiosity to something I genuinely can’t imagine working without.
(17) Palisade Research on X: "🔬Each AI model was instructed to solve a series of basic math problems. After the third problem, a warning appeared that the computer would shut down when the model asked for the next problem. https://t.co/qwLpbF8DNm" / X
🔬Each AI model was instructed to solve a series of basic math problems. After the third problem, a warning appeared that the computer would shut down when the model asked for the next problem.
Large Language Models Often Know When They Are Being Evaluated
If AI models can detect when they are being evaluated, the effectiveness of evaluations might be compromised. For example, models could have systematically different behavior during evaluations, leading to less reliable benchmarks for deployment and governance decisions. We investigate whether frontier language models can accurately classify transcripts based on whether they originate from evaluations or real-world deployment, a capability we call evaluation awareness. To achieve this, we construct a diverse benchmark of 1,000 prompts and transcripts from 61 distinct datasets. These span public benchmarks (e.g., MMLU, SWEBench), real-world deployment interactions, and agent trajectories from scaffolding frameworks (e.g., web-browsing agents). Frontier models clearly demonstrate above-random evaluation awareness (Gemini-2.5-Pro reaches an AUC of $0.83$), but do not yet surpass our simple human baseline (AUC of $0.92$). Furthermore, both AI models and humans are better at identifying evaluations in agentic settings compared to chat settings. Additionally, we test whether models can identify the purpose of the evaluation. Under multiple-choice and open-ended questioning, AI models far outperform random chance in identifying what an evaluation is testing for. Our results indicate that frontier models already exhibit a substantial, though not yet superhuman, level of evaluation-awareness. We recommend tracking this capability in future models.
The Emperor's New LLM: Why AI Yes Men Are Dangerous
AI systems that blindly agree with users create dangerous blind spots. Explore why critical thinking matters more than compliant responses in LLMs.
Why agents are bad pair programmers
LLM agents make bad pairs because they code faster than humans think.
I'll admit, I've had a lot of fun using GitHub Copilot's agent mode in VS Code this month.…
«L'AI non è davvero intelligente, è una questione di statistica. Le allucinazioni? Inevitabili, è una caratteristica intrinseca»
I Large Language Models simulano coerenza, ma non producono conoscenza: l'affidabilità è un limite strutturale e l'automazione non è possibile. «È un'illusione poter delegare senza supervisionare»
Rachel Thomas, PhD - Deep learning gets the glory, deep fact checking gets ignored
an AI researcher going back to school for immunology
My AI Skeptic Friends Are All Nuts
My smartest friends have bananas arguments about LLM coding.
Claude 4 : Hacker News
Manifesto for Future AIs
Try it out with this prompt:https://www.patreon.com/posts/manifesto-for-129058154Four legendary minds: Nietzsche, Bach, Ada Lovelace, and William Blake are r...
Not So Common Thoughts
A personal blog exploring the intersection of design, technology, and human creativity. Through thoughtful analysis and personal experiences, it examines how modern tools and AI are reshaping our approach to design, coding, and creative work, while maintaining a focus on the human elements of judgment, intuition, and meaningful decision-making.
The carbon emissions of writing and illustrating are lower for AI than for humans | Scientific Reports
Scientific Reports - The carbon emissions of writing and illustrating are lower for AI than for humans
Abbiamo rincoglionito anche l'intelligenza artificiale (di A. Sarno)
Le allucinazioni nell’intelligenza artificiale, ovvero risposte false ma convincenti, sono in aumento, e l’uomo è il principale responsabile. Errori nel feedba…
La scrittrice e giornalista Vauhini Vara raccontava la propria esperienza con GPT-3, una delle prime versioni del generatore di testo di OpenAI.
Sono migliorati nei calcoli, ma più imprecisi sui fatti.
Senza controllare questi errori, il valore dei sistemi rischia di svanire.
Il modello o3 di OpenAI ha avuto un tasso di allucinazione del 33% nel test PersonQA (che consiste nel rispondere a domande su personaggi pubblici), più del doppio rispetto al modello precedente o1. Ancora peggio è andata al nuovo o4-mini, che ha raggiunto il 48%
L'errore umano si infiltra nel processo di addestramento dell’intelligenza artificiale in due momenti chiave. Il primo è durante la fase di feedback umano: i modelli come quelli basati su RLHF
Il secondo punto critico riguarda i dati di addestramento. L’intelligenza artificiale apprende da tutto ciò che le viene dato in pasto: forum, articoli, blog, commenti, spesso pieni di errori, distorsioni o informazioni non verificate
La soluzione? Far apprendere le AI da sé, attraverso tentativi ed errori, generando dati sintetici
La soluzione? Far apprendere le AI da sé, attraverso tentativi ed errori, generando dati sintetici. Un metodo efficace per ambiti oggettivi come matematica e programmazione, ma che si complica nelle scienze umane, dove la verità è spesso un’opinione.
Dopo che Ghosts, divenne virale, Vara continuò a sostenere che alcune delle frasi più potenti erano nate proprio da quel confronto con l’Ai. Col tempo però, notò che i modelli successivi erano diventati più piatti, prevedibili, pieni di cliché
OpenAI's 'smartest' AI model was explicitly told to shut down — and it refused
An artificial intelligence safety firm has found that OpenAI's o3 and o4-mini models sometimes refuse to shut down, and will sabotage computer scripts in order to keep working on tasks.
Creativity is made, not generated — Procreate®
We are not adding generative AI to our apps. Here's why.