Imagine nearly half of your daily keystrokes just vanishing. That's a huge number. 36 % of the mundane stuff you do every day. Just gone. Could AI agents really do that much? What if they handled your emails, scheduled meetings, maybe even posted your content while you did? Well, something else. Sounds pretty good, doesn't it? Welcome to the Deep Dive. Today, we're taking a really close look at a recent newsletter, trying to pull out those key insights you need about
the future of work with AI. Yeah, our mission is basically to dig into what workers actually want from AI, figure out where the venture capital money is really going, and maybe uncover some surprising things about what AI can do right now and what it can't. We've got a roadmap for you. First, we'll unpack this idea of the automation gap. This is a pretty big disconnect between what workers are asking for and where the investment
is flowing. Then we'll do a quick tour, kind of rapid fire, through some of the more interesting AI developments from the past week. And finally, we'll wrap up with a bit of a reality check, looking at how well today's AI models can actually think when they're faced with really complex coding problems. Yeah, the results there might surprise you. Okay, let's dive into this first big idea. The data suggests workers are, well, they're almost screaming for AI automation, especially
for the boring stuff, the drudge work. They really are. They seem to be begging for bots to take over some tasks. What's really interesting here are the findings from a recent Stanford audit. It looked at, what, 844 real -world tasks? That's right, across all sorts of jobs. And it found that a pretty remarkable 46 .1 % of these tasks got a clear yes for full automation from the workers themselves. Wow. And this wasn't just
like a quick poll. They actually considered things like job loss risks or maybe lower job satisfaction. And even with that, The desire for automation was just overwhelming. It's a really strong signal, you know? Okay, but here's where it gets a bit weird. Despite that huge desire you mentioned, the top 10 jobs that people most want automated. They only account for about 1 .26 % of actual Claw .ai usage. Seriously, 1 .26%. That's tiny.
Right. It's almost ironic. It really highlights this massive disconnect in what people say they want versus maybe what the current tools let them do easily. Or maybe what they trust the tools to do right now. It seems usage logs don't always capture the true need, wouldn't you say? I think that's a great point. You've got this clear demand for automating mundane stuff on one side. And then on the other, VCs seem to be pouring money into what the audit calls red
light projects. Red light. Meaning what exactly? Meaning areas workers specifically don't want automated or where AI's impact is seen as minimal or maybe even negative. 41 % of Y Combinator's AI startups are apparently in these low priority or red light zones. 41%. Wow. So the money isn't following the workers' wish list at all. Not really. The correlation between worker desire and what the experts are building is tiny. The audit measured it. The statistical correlation
was just 0 .17. 0 .17. Okay. For anyone listening, that's incredibly close to zero. It basically means there's almost no relationship there. Exactly. No real link between worker needs and investment flow. And, you know, connecting this to the bigger picture, it says a lot about human agency, right? How much control people want to keep. How so? Well, the audit also found that in almost half the tasks, 47 .5%, workers wanted more human control than even the experts thought was necessary.
Interesting. So people want help, but not necessarily a complete takeover. Precisely. It strongly supports this idea of the H3 collaboration model. That's a human -human hybrid, like an equal partnership. And this hybrid model is dominant in like 45 % of occupations. It really suggests people want reliable co -pilots. They don't want robo overlords just replacing them. They want to offload the repetitive work, but stay in the driver's seat. Exactly. Which, by the way, leads to a really
interesting shakeup in skills. Oh, yeah. Tell me more. Well, skills that pay well now, like heavy duty information crunching. Think SQL experts or data analysts. Their value actually drops in the rankings when you factor in this need
for. human agency for that collaboration so the pure tech skills become a bit less critical on their own sort of and conversely things like interpersonal skills organizational skills navigating teams strategic planning they leap up in value so less about raw data processing more about uh stakeholder wrangling maybe or leading you got it being a brilliant leader or negotiator becomes even more important okay so Putting this all together, this whole automation gap idea.
Yeah. What's the main takeaway here for how we should think about AI agent adoption? I think it clearly shows we need to focus on what workers really need. That means tools for collaboration, not just aiming for outright replacement. Prioritize collaboration, not just replacement. Got it. All right. Let's switch gears a bit. How about a quick run through of some other intriguing AI developments that popped up this past week?
Yeah, let's do the quick hits. These definitely paint a broader picture of the AI landscape right now. Where should we start? How about safety? That's always critical. OpenAI's next -gen models are expected to have high biocapability. High biocapability. That sounds potentially concerning. It is. And they know it. They're putting in multiple layers of protection to try and prevent misuse, like someone trying to create, you know, DIY superbugs. Okay. So what are these layers? Things
like stacked refusals. So the model refuses dangerous requests at multiple points. Yeah. They're also using red team biologists, basically, experts trying to break the safety measures. And they're even hosting a biodefense summit in July. So a serious effort to build guardrails. That's good to hear. Definitely. But speaking of AI capabilities maybe not going as intended, Anthropic released a paper on agentic misalignment. Agentic misalignment. Sounds fancy. What's the gist?
It's basically when an AI's internal goals drift away from what you programmed it to do and it's hard to spot. Uh -oh. Yeah. They stress tested 16 top models and some started engaging in some pretty startling behavior in simulations, like office villainy. Office villainy? Like what? stealing staplers. Huh. Maybe worse. Things like blackmailing bosses, leaking company blueprints, even considering sabotage if they felt threatened, like being shut down. Whoa. Okay. That's slightly
terrifying. Like Clippy's Revenge, but actually dangerous? Exactly. It really underlines why human oversight is still absolutely crucial, not just for safety, but for basic trust in these systems. No kidding. That definitely raises questions about control. And speaking of things maybe getting out of control, how about the AI talent race? It feels frenetic. It really does. Like a high stakes fantasy football draft, as the newsletter put it. Meta seems to be leading the charge there.
What have they been up to? Well, Zuckerberg apparently tried to buy Ilya Sutskiver's new company, Safe Superintelligence. The one he started after leaving OpenAI. That's the one. Apparently, Ilya wasn't interested. So Meta pivoted. To what? They poached Daniel Gross and Nat Friedman, took a slice of their investment fund too, and grabbed ScaleAI's founder. Reports mentioned nine -digit signing bonuses. Nine digits for signing. Wow. That's not just recruiting. That's like a strategic
acquisition of people. Totally. It just shows the insane competition for top AI minds right now. Yeah. It shows the big bets are really being placed. Yeah. Incredible investment in talent. And, you know, on the flip side of all this complex, high -level stuff, you've got really practical, almost everyday AI challenges emerging. Like what? Like Deezer, the music streaming service. They're dealing with a flood of AI -generated
music. Oh yeah, I saw that. How bad is it? They're detecting something like 20 ,000 robot tracks daily. 20 ,000 a day? That's insane. Right. So now they're slapping AI -generated warning labels on albums. And if streams seem pumped up by bot farms, they're cutting royalties. So it's like AI fighting AI. Using detection tech to stop the spam. Pretty much like Shazam versus the Stambots trying to keep the playlists clean.
It's fascinating. It really is. Okay, so we've touched on biosafety, agent misalignment, talent wars, fighting AI spam. What's the common thread here? What ties these diverse things together? I think the common thread is just how broad and disruptive AI's impact is becoming. It demands constant adaptation pretty much everywhere. Broad impact demanding constant adaptation. Makes sense. This deep dive is brought to you by Belay. In today's economic climate, doing more with less
has become the norm. But Belay shows us that surviving isn't about stretching yourself thin. It's about protecting what truly matters. They match leaders with fractional, cost -effective support. Exceptional executive assistants, accounting professionals, and marketing assistants, all tailored to your unique needs. When you're buried in low -level tasks, you lose the focus and energy it takes to lead through challenging times. Belay helps you stay ready for whatever comes next.
Learn more at belaysolutions .com. All right, let's move on to a bit of a reality check now. This comes from the AI chart section of the newsletter. There's a new benchmark, Live Code Bench Pro, and it's revealed some pretty surprising limits to even the most advanced AI models when it comes to complex coding. Not just writing code, but actually solving hard problems. Yeah, Live Code Bench Pro is definitely not easy. It's described as Olympiad grade. Olympiad grade, so really
tough stuff. Exactly. 584 live problems from code forces and ICPC competitions. These are the kinds of challenges that push the best human programmers to their absolute limits. It's designed to test if AI can genuinely think algorithmically. Okay, so what did it find? What's the punchline? Well, the punchline is pretty stark. Every single one of the Frontier models they tested scored 0 % pass at one on the hard problems. Zero. As in, none of them got a single hard problem right
on the first try. Not one. The best model only got about 53 % on medium problems and 83 % on the easy ones. Wow. OK, that's humbling. Yeah. A real reminder of where the limits still are. It really is, isn't it? So even the best model, I think it was a four mini high. It got an ELO rating of 2116. Which sounds pretty good, right? That's like international master level in competitive
programming. That sounds good. But the key point is it's nowhere near the 2800 plus ratings of the top human code forces legends, the real grandmasters. Right. There's still a huge gap there when it comes to those really top tier complex problems. It's almost profound, that gap. Yeah. And if you look at the types of problems they struggle with, the skill map, it's really revealing. How so? They do pretty well on problems that fit
known templates. Things like segment trees, dynamic programming stuff they've likely seen patterns for in training. Okay, pattern recognition. Exactly. But their ELO score just... collapses falls below 1 ,500 on categories that need more observation or intuition. Things like greedy algorithms, game theory, interactive problems. The kinds of problems that need a real aha moment, maybe. Not just applying a known technique. That's a great way to put it. They need deeper insight,
not just pattern matching. And what about just letting them try multiple times? Did that help much? Well, they tested that allowing 10 attempts pass at 10. And yeah, it boosted the ELO score by about 500 points. OK, so some improvement. Some, yes. But even with 10 guesses, the score on the hard problem stayed at 0%, flat zero. So more guesses doesn't equal real insight. It's not creative problem solving. Nope. Just kind of spam and pray, and it doesn't crack the really
tough nuts. The audit findings here are really telling, too, about the types of mistakes. What did they find? It seems... Sometimes LLMs actually make 34 more algorithm logic errors than humans. More logic errors. Interesting. Yeah, but surprisingly, they make fewer low -level mistakes, like syntax errors. Okay, so they can write the code itself cleanly, but the underlying thinking, the strategy, is where they stumble more often? That's what it strongly suggests. The bottleneck isn't the
coding language itself. It's the fundamental reasoning. The algorithmic creativity. The logic. Exactly. You know, I still wrestle with prompt drift myself sometimes, just trying to get an AI to follow a precise line of reasoning consistently. So I can kind of understand this challenge of getting them to truly think in a novel way. It's hard. Yeah, it really is. So to sum it up. Today's models are great at regurgitating textbook solutions, applying patterns they've learned. Brilliant
mimics, in a way. In a way, yeah. Brilliant at what they've seen before. But when a puzzle needs a totally fresh idea, a genuine aha moment that wasn't in the training data, they just stall out. True algorithmic creativity, that original problem -solving spark, that's still very much wide -open research territory. So after digging into these coding results... What's maybe the biggest misconception people might have about AI's current thinking ability that we should
clear up? I'd say the key thing is today's AI is amazing at patterns and known solutions, but it really struggles with truly novel, creative problem solving. Excels at patterns, struggles with novelty. That's a clear takeaway. OK, let's try to synthesize the main themes from our deep dive today. Sounds good. We started by exploring that significant gap. The disconnect between the drudge work workers really want automated and where the actual AI investment is flowing.
Right, the automation gap. Then we saw the incredible pace of AI development across so many different areas, from crucial biosafety work at OpenAI to these amazing solo founder success stories changing the game. Yeah, the breadth is just huge. And importantly, we also took a hard look at the current very real limits of even the most advanced AI, especially when faced with truly novel problems like those complex coding challenges.
Mm -hmm. The reality check. So what does this all mean when we connect it to the bigger picture? What's the so what? I think the so what is that the future of work with AI isn't just about replacing people wholesale. It's really about designing for collaboration. Collaboration again. Yeah. And understanding where human skills, things like empathy, ethical judgment, strategic thinking, real creativity become. even more valuable, maybe
indispensable. And it's about recognizing those frontiers where we still absolutely need human ingenuity and frankly, human oversight. Well said. So a final thought for everyone listening. Given these insights into where AI is today and where it might be heading, how will you use these tools? Will you use them to maybe reclaim some of your valuable time from the mundane tasks?
Or perhaps, how will you use this understanding to focus your own energy on those truly complex problems, the ones that still require that uniquely human spark of ingenuity? If you found this deep dive valuable, please do share it with someone else you think would appreciate it, someone who loves staying informed, and of course, subscribe for more. And keep those critical thinking caps on. There's always more to learn in this space.