What are the key takeaways from this Deep Questions with Cal Newport episode?

Key insights include: **Scaling Law Failure:** The 2020 Kaplan paper showed language models improved dramatically when made bigger, enabling GPT-3 and GPT-4 breakthroughs. By fall 2024, this stopped working—OpenAI's Orion, Meta's Behemoth, and Elon Musk's Grok-3 all failed to deliver expected leaps despite massive compute investments, ending the path to AGI.; **Post-Training Shift:** After pre-training scaling failed, AI companies pivoted to post-training techniques like reinforcement learning and test-time compute to squeeze better performance from existing models. This produced incremental improvements measured by benchmark percentages rather than transformative new capabilities, fundamentally changing the industry trajectory from revolutionary to evolutionary progress.; **Job Market Reality:** Media reports conflate unrelated factors—tech sector layoffs stem from pandemic overhiring corrections, not AI replacement. A resurfaced MIT study found 95% of companies attempting AI implementation failed and abandoned it. Actual AI revenue totals only 35 billion dollars annually versus 560 billion in capital expenditures over eighteen months.

What did Dario Amede and Ed Zittron discuss on Deep Questions with Cal Newport?

Cal Newport examines why GPT-5 disappointed expectations, revealing how AI scaling laws stopped working in 2024, forcing companies to shift from breakthrough pre-training to incremental post-training improvements while overstating economic disruption claims. Key topics include: **Scaling Law Failure:** The 2020 Kaplan paper showed language models improved dramatically when made bigger, enabling GPT-3 and GPT-4 breakthroughs. By fall 2024, this stopped working—OpenAI's Orion, Meta's Behemoth, and Elon Musk's Grok-3 all failed to deliver expected leaps despite massive compute investments, ending the path to AGI.; **Post-Training Shift:** After pre-training scaling failed, AI companies pivoted to post-training techniques like reinforcement learning and test-time compute to squeeze better performance from existing models. This produced incremental improvements measured by benchmark percentages rather than transformative new capabilities, fundamentally changing the industry trajectory from revolutionary to evolutionary progress..

How long is this episode of Deep Questions with Cal Newport?

This episode is 97 minutes long. SignalCast provides an AI-generated summary so you can get the key insights in about 3 minutes.

Deep Questions with Cal Newport

Ep. 367: What if AI Doesn’t Get Much Better Than This?

August 25, 2025

97 min episode · 2 min read

Dario Amede,Ed Zittron

Episode

97 min

Read time

2 min

Topics

Career Growth, Investing, Startups

AI-Generated Summary

Published Dec 25, 2025

Key Takeaways

✓Scaling Law Failure: The 2020 Kaplan paper showed language models improved dramatically when made bigger, enabling GPT-3 and GPT-4 breakthroughs. By fall 2024, this stopped working—OpenAI's Orion, Meta's Behemoth, and Elon Musk's Grok-3 all failed to deliver expected leaps despite massive compute investments, ending the path to AGI.
✓Post-Training Shift: After pre-training scaling failed, AI companies pivoted to post-training techniques like reinforcement learning and test-time compute to squeeze better performance from existing models. This produced incremental improvements measured by benchmark percentages rather than transformative new capabilities, fundamentally changing the industry trajectory from revolutionary to evolutionary progress.
✓Job Market Reality: Media reports conflate unrelated factors—tech sector layoffs stem from pandemic overhiring corrections, not AI replacement. A resurfaced MIT study found 95% of companies attempting AI implementation failed and abandoned it. Actual AI revenue totals only 35 billion dollars annually versus 560 billion in capital expenditures over eighteen months.
✓Computer Science Careers: Master's degrees in computer science typically provide positive salary returns because two-year programs enable higher starting positions that offset lost earnings. PhDs require five-plus years and should only be pursued for research careers, not pure salary optimization. Degree quality and institutional reputation matter significantly for hiring outcomes.
✓Digital Minimalism Practice: Ed Sheeran eliminated his phone in 2015 after accumulating 10,000 contacts, switching to iPad-only email checked weekly. People adapted without conflict—no enforcement mechanisms exist requiring instant availability. The feared social consequences of communication boundaries rarely materialize; others simply adjust their expectations and move forward with their lives.

What It Covers

Cal Newport examines why GPT-5 disappointed expectations, revealing how AI scaling laws stopped working in 2024, forcing companies to shift from breakthrough pre-training to incremental post-training improvements while overstating economic disruption claims.

Key Questions Answered

•Scaling Law Failure: The 2020 Kaplan paper showed language models improved dramatically when made bigger, enabling GPT-3 and GPT-4 breakthroughs. By fall 2024, this stopped working—OpenAI's Orion, Meta's Behemoth, and Elon Musk's Grok-3 all failed to deliver expected leaps despite massive compute investments, ending the path to AGI.
•Post-Training Shift: After pre-training scaling failed, AI companies pivoted to post-training techniques like reinforcement learning and test-time compute to squeeze better performance from existing models. This produced incremental improvements measured by benchmark percentages rather than transformative new capabilities, fundamentally changing the industry trajectory from revolutionary to evolutionary progress.
•Job Market Reality: Media reports conflate unrelated factors—tech sector layoffs stem from pandemic overhiring corrections, not AI replacement. A resurfaced MIT study found 95% of companies attempting AI implementation failed and abandoned it. Actual AI revenue totals only 35 billion dollars annually versus 560 billion in capital expenditures over eighteen months.
•Computer Science Careers: Master's degrees in computer science typically provide positive salary returns because two-year programs enable higher starting positions that offset lost earnings. PhDs require five-plus years and should only be pursued for research careers, not pure salary optimization. Degree quality and institutional reputation matter significantly for hiring outcomes.
•Digital Minimalism Practice: Ed Sheeran eliminated his phone in 2015 after accumulating 10,000 contacts, switching to iPad-only email checked weekly. People adapted without conflict—no enforcement mechanisms exist requiring instant availability. The feared social consequences of communication boundaries rarely materialize; others simply adjust their expectations and move forward with their lives.

Notable Moment

Newport reveals that by summer 2024, all major AI companies privately knew their scaling strategies had failed. OpenAI's GPT-5 used five to ten times more compute than GPT-4 but delivered only marginal improvements, while tech CEOs continued making grandiose AGI claims publicly despite internal disappointments.

Know someone who'd find this useful?

Episode Transcript

In the years since ChatGPT's astonishing launch, it's been hard not to get swept up in feelings of euphoria or dread about the looming impacts of this new type of artificial intelligence. But in recent weeks, this vibe seems to be shifting. Both the media and technologists no longer seem so certain that everything is about to change. Now how is this possible? What went wrong? What should we really expect from this tech in the next few years? I'm Cal Newport, and this is Deep Questions. Today's episode, what if AI doesn't get much better than this? Part one, the week we woke up. Dario, you've said that AI could wipe out half of all entry level white collar jobs and spike unemployment to 10 to 20%. How soon might that happen? Well, let's, well, first of all, thanks for having me on the show. But, just to back up a little bit, you know, I've been building AI for over a decade. And I think maybe the most salient feature of the technology and what is driving all of this is how fast the technology is getting better. A couple years ago, you could say that AI models were maybe as good as a smart high school student. I would say that now they're as good as a smart college student and and and sort of reaching past that. I really worry particularly at the entry level that the AI models are are are are, you know, very much at the center of what what an entry level human worker would do. That was Dario Amede talking to CNN's Anderson Cooper. Now Dario is the CEO of the AI company Anthropic. And if you wanna know why we have become so worked up about generative AI, a big part of this answer is that tech CEOs like Amade have been saying astonishing claims like the one we just heard. Remember what he just said there. AI used to be as good as a average high school student. Now they're as good as a smart college student, and he worries about entry level jobs still being around for humans to actually do. Now Amadeh is not alone in these types of claims. If If you listen to the CEOs of these companies in the last six months or so, it's almost like they've had a competition to see who could become more over the top. Jesse, play the clip we have of Sam Altman, CEO of OpenAI, appearing on Theo Vaughn's podcast. You know, there there are these moments in the history of science where you have a group of scientists look at their creation and just say, you know, what what have what have we done? What maybe it's great. Maybe it's bad. But what have we done? Like, maybe the most iconic example is thinking about the scientists working on the Manhattan Project in 1945, sitting there watching the Trinity test and just, you know, this thing …

Get the full transcript (19,121 words) + summary by email — free

One-time email with the complete transcript and AI summary of this episode. No account needed.

One email, no spam. We’ll also show you what SignalCast does.

Browse all Deep Questions with Cal Newport transcripts →