weekend ai reads for 2025-06-13

By comparing LRMs with their standard LLM counterparts under equivalent inference compute, we identify three performance regimes: (1) low-complexity tasks where standard models surprisingly outperform LRMs, (2) medium-complexity tasks where additional thinking in LRMs demonstrates advantage, and (3) high-complexity tasks where both models experience complete collapse. We found that LRMs have limitations in exact computation: they fail to use explicit algorithms and reason inconsistently across puzzles. We also investigate the reasoning traces in more depth, studying the patterns of explored solutions and analyzing the models’ computational behavior, shedding light on their strengths, limitations, and ultimately raising crucial questions about their true reasoning capabilities.

Worse still, the models don’t seem to know they’re failing. They produce what appear to be sound, step-by-step answers. To a lay reader, or even a trained one, these outputs can seem rational. But they’re not grounded in any algorithmic or consistent method. They’re approximations of logic based on semantic coherence.

Apple did this by showing that leading models such as ChatGPT, Claude and Deepseek may “look smart – but when complexity rises, they collapse”. In short, these models are very good at a kind of pattern recognition, but often fail when they encounter novelty that forces them beyond the limits of their training, despite being, as the paper notes, “explicitly designed for reasoning tasks”.

It seems important that this doesn’t follow?

1. Not doing [X] in a given situation doesn’t mean you can’t do [X] in general.
2. Not doing [X] in a particular test especially doesn’t mean a model can’t do [X].
3. Not doing [X] can be a simple ‘you did not provide enough tokens to [X]’ issue.
4. The more adversarial the example, the less evidence this provided.
5. Failure to do any given task requiring [X] does not mean you can’t [X] in general.

Or more generally, ‘won’t’ or ‘doesn’t’ [X] does not show ‘can’t’ [X]. It is of course often evidence, since doing [X] does prove you can [X]. How much evidence it provides depends on the circumstances.

Tasted a little tear gas— tasted like fascism

Like Dario Amodei’s Machines of Loving Grace, the latest Altman essay spends the bulk of its time hand waving away all the important concerns about such futures, both in terms of getting there and where the there is we even want to get. It’s basically wish fulfillment.

AI EXPERTS: 1% of the workforce

Al experts are the most proficient Al users in the workforce, as well as the most bullish about Al’s potential. They receive the most resources for Al development, and 31% of them are saving more than 12 hours - or a day and a half - of their work week by using Al.

Their stats

• Most likely to be C-suite

• 84 average proficiency score out of 100

• 69% are daily Al users

• 47% are saving 8-12 hours per week using Al

Their differentiators

• 94% are in companies that approve of Al

• 71% have received Al training

• 84% have managers that encourage Al

• 75% report their company having a clear Al policy

Just a few years ago, lip-syncing a minute of animation could take up to four hours. An animator would listen to an audio track and laboriously adjust character mouths frame by frame. But Mr. Peck’s one-minute scene took 15 minutes for the A.I. tool to sync, including time spent by an artist to refine a few spots by hand.

As agents take on high-frequency or transactional work, employees shift into roles that require more oversight, ethics, and judgment, including:

- Custodians whoensure the integrity of data, model performance, and customer outcomes.
- Judgment holders who handle ambiguous or high-stakes decisions where context, nuance, and trust are essential.
- Approvers and auditors who review exceptions, manage escalations, and reinforce compliance boundaries.

First, configuration is unintuitive. MCPs work like IFTTT where you need to establish connections on both the app side and the LLM side to make them function. This creates setup friction that most users won’t tolerate.

Second, the UX approach doesn’t feel right for the future of apps with natural language capabilities. The way MCPs bolt conversational AI onto existing tools feels like a bridge solution rather than how apps will naturally evolve. The interaction patterns aren’t optimized for mass-market usage.

To understand why these limitations matter, it’s important to see the types of workflows that MCPs help with today.

Instead, leaders should set their teams up for success by helping to identify opportunities for AI-human collaboration. One way to do this is for teams to ensure that agents are working with good data, weeding out wrong or incomplete information and regularly optimizing and providing feedback.

In my experience, these organizations aren’t necessarily resistant to innovation. Instead, they’re working to balance progress with the need for stability, compliance, and alignment with existing operations. In these cases, appointing a Chief AI Officer can serve a valuable purpose by creating a focal point for AI strategy and helping to coordinate efforts across departments.

Speed isn’t just important, it is the moat. The ability to build, ship, learn, and adapt faster than everyone else is the only sustainable edge right now. In a world where everything is open source, everything is demo-able, and everything is one blog post away from being copied, speed is the only thing that compounds.

Part of our task in the face of generative AI is to make an argument for the value of thinking – laboured, painful, frustrating thinking. It is not an easy sell. But to give up on this is to give up on our students, most of whom are at an age where they can be easily seduced by techno-sirens promising instantaneous essays for minimal effort and with little chance of getting caught. They deserve better from us.

While many of us are aware we should be mindful about uploading sensitive documents to AI systems, talking to a bot like it is a person and habitually revealing personal information to it is an extraordinary security risk when you deal with sensitive data. Our words are now prompts, our conversations become data, and the potential FERPA and HIPPA violations that may come from talking about someone with something is not being discussed enough.

The PR worker also didn’t seem to be doing “higher-level work,” but simply doing analysis more quickly. The output provided by AI is clearly useful to a junior worker’s bosses, but I’m skeptical that it’s giving them a deeper understanding of how a business or industry works.

TL;DR: I run Claude Code in no-prompt mode; it saves me an hour a day and hasn’t broken my Mac in two months. The $200/month Max plan pays for itself.

For the past two months, I’ve been living dangerously. I launch Claude Code (released in late February) with --dangerously-skip-permissions, the flag that bypasses all permission prompts. According to Anthropic’s docs, this is meant “only for Docker containers with no internet”, yet it runs perfectly on regular macOS.

SeventhSight app uses patented Machine Learning Artificial Intelligence to analyse the meaning of your dream. Find powerful insights into your daily life by understanding what your subconscious is telling you!

Timbaland has launched his own AI entertainment company called Stage Zero and its first signee is the artist TaTa. Co-founded with Rocky Mudaliar and Zayd Portillo, Stage Zero’s first signee is an AI pop artist called TaTa, driven by Suno AI. The pop artist, along with a bevy of AI-driven creative tools will all be under Timbo’s new company.

Mental Health Tip: Do this research using voice mode while walking. Legal research isn’t fun, but being outside in fresh air using your voice rather than staring at a screen makes it much more bearable.