weekend ai reads for 2025-05-09

Having read his chat logs, she only found that the AI was “talking to him as if he is the next messiah.” The replies to her story were full of similar anecdotes about loved ones suddenly falling down rabbit holes of spiritual mania, supernatural delusion, and arcane prophecy — all of it fueled by AI. Some came to believe they had been chosen for a sacred mission of revelation, others that they had conjured true sentience from the software.

The best-performing model was Anthropic’s Claude 3.5 Sonnet, which struggled to finish just 24 percent of the jobs assigned to it. The study’s authors note that even this meager performance is prohibitively expensive, averaging nearly 30 steps and a cost of over $6 per task.

All the time we’re agonizing about our choices, we’re avoiding the hard work of really considering how we sorted things in the first place.

If there is a difference, it lies between the text that seeks to produce a new reader and the text that tries to fulfill the wishes of the readers already to be found in the street. In the latter case we have the book written, constructed, according to an effective, mass-production formula; the author carries out a kind of market analysis and adapts his work to its results. Even from a distance, it is clear that he is working by a formula; you have only to analyze the various novels he has written and you note that in all of them, after changing names, places, distinguishing features, he has told the same story —the one that the public was already asking of him.

Nonetheless, the answer to the question “What customer needs requires an AI solution?” still isn’t always “yes.” Large language models (LLMs) can still be prohibitively expensive for some, and as with all ML models, LLMs are not always accurate. There will always be use cases where leveraging an ML implementation is not the right path forward.

The state asked for a 9.5-year sentence, and the judge ended up giving Horcasitas 10.5 years for manslaughter, after being so moved by the powerful video, family says. The judge even referred to the video in his closing sentencing statements.

But in addition to the utility of the app, MenuGen was interesting to me as an exploration of vibe coding apps and how feasible it is today. As such, I did not write any code directly; 100% of the code was written by Cursor+Claude and I basically don't really know how MenuGen works in the conventional sense that I am used to.

Google/Gemini

Fastest top-tier models. Huge context window. Good prices.

Issue 1: Setup

The setup is SO bad. The main API doesn't follow the spec. They have an openai-compat mode but it sucks.

Vertex is SO bad. It's the only major AI platform that doesn't let you use traditional API keys. It's built to be plugged into the rest of the GCP ecosystem. Sadly, that ecosystem is really rough and most AI apps are not part of it.

As such, we use AI Studio. Much better! But also much more limited.

Having the best individual features is not what makes compound software special. Rather the additional product-market fit that it unlocks by having multiple products/features across the company. This is where the magic is created.

Despite finding widespread and often employer-encouraged adoption of these tools, the study concluded that “AI chatbots have had no significant impact on earnings or recorded hours in any occupation” during the period studied. The confidence intervals in their statistical analysis ruled out average effects larger than 1 percent.

That future may arrive sooner than expected when you consider what a short window college really is. Already, roughly half of all undergrads have never experienced college without easy access to generative AI.

Much like pencil and calculators, Ingram acknowledged AI is “here to stay,” but said it can only reach its full potential under the guidance of trained educators who know how best to integrate the technology into their classrooms.

We highlight how AI, trained on biased human data, can perpetuate societal inequities and discriminatory practices despite appearing objective. We present a technology audit that examines how LLMs score and provide feedback on student writing samples paired with student descriptions. Findings reveal that LLMs exhibit implicit biases, such as assigning lower scores when students are said to attend an “inner-city school” or prefer rap music.

NeMo helps enterprise AI developers easily curate data at scale, customize large language models (LLMs) with popular fine-tuning techniques, consistently evaluate models on industry and custom benchmarks, and guardrail them for appropriate and grounded outputs.

During inference, the user will provide the LLM with a text, and the LLM, based on the probabilities it’s learned through training, will decide what token comes next. However, it will not decide just one token: it will take into consideration every possible token that exists in its vocabulary, assigns a probability score to each, and (depending on your sampler) will only output the most probable token, i.e. the one with the highest score. This would make for a rather boring output (unless you need determinism), so this is where Sampling comes in.

Imagine a future where everyone is a lifelong learner, supported in their wellbeing, and empowered to not just advance in their careers but to flourish as human beings.

I think what’s happened with AI is, instead of the gains being that I’m faster, it’s that the scope of my projects has gotten bigger.

The evolution of these metrics is important to understanding the all-consuming nature of surveillance capitalism — even a scathing rant about surveillance capitalism becomes fodder for the machine, as you can clearly see with the ads on this page. When deep-pocketed advertisers are involved, positive and negative traits alike become dollar signs; search terms to be probed, analyzed, and used for profit.