September 8, 2025 • 21 mins
Apple to Revamp Siri and Search with AI-Powered Tool by 2026
OpenAI to Develop In-House AI Chip with Broadcom to Reduce Nvidia Dependence
AI's Potential Unfulfilled: Only 5% of Projects Add Value, Operational Gains Key
Mistral AI Secures €2 Billion Investment, Valued at $14 Billion
Meta Considers Partnerships with Google or OpenAI to Boost AI in Apps
Meta Restructures Leadership to Dominate AI Sector, Focuses on Strategic 2026 Growth
Microsoft Tests In-House AI Model MAI-1 to Compete with OpenAI
Nvidia reports record sales as the AI boom continues
The GPT-5 Disappointment
#AI, #OpenAI, #Siri, #Meta, #Nvidia, #Microsoft, #Broadcom
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
(00:00):
Welcome to Innovation Pulse, your quick no-nonsense update on the latest in AI.
(00:09):
First, we will cover the latest news.
Apple is revamping Siri with AI.
Open AI is launching its first AI chip, and Mistral AI is securing major investment.
After this, we'll dive deep into the controversial launch of GPT-5 by Open AI.
Apple is set to revamp Siri, Apple Intelligence, and Apple Search for 2026.
(00:33):
Bloomberg reports that Apple plans to launch an AI-powered web search tool integrated into
Siri.
Working alongside Google, the new system, named World Knowledge, answers, will search the
Internet and provide AI-driven summaries for easier and more accurate results.
This feature will also be integrated into Safari, Spotlight, and iPhone's default search.
(00:59):
The technology behind this, called Linwood and LLM Siri, is part of a software update
internally known as LUCK-E, possibly arriving in iOS version 26.4 by March.
Apple is collaborating with Google to test an AI model to enhance the voice assistant
with the new search, including text, photos, video, and local information.
(01:23):
The impact on SEO remains to be seen.
Open AI is preparing to launch its first AI chip next year in partnership with Broadcom,
according to the Financial Times.
The chip is intended for internal use rather than for external customers.
This move aligns with efforts by tech giants like Google, Amazon, and Meta, who have developed
(01:50):
custom chips to manage AI workloads.
Open AI aims to diversify its chip supply and reduce costs, previously collaborating with
Broadcom and Taiwan semiconductor manufacturing co, alongside using AMD and Nvidia chips.
In February, it was reported that Open AI was advancing its plan to lessen dependence
(02:12):
on Nvidia by developing its in-house AI silicon.
Broadcom CEO Hock Tan mentioned expected significant AI revenue growth by 2026, with over $10 billion
in AI infrastructure, orders secured from a new customer.
This strategic shift reflects the growing demand for computing power to train and operate
(02:36):
AI models.
Up next, we're exploring AI's impact on operations.
MIT's recent study brings surprising insights for those anxious about AI replacing jobs.
Although many companies attempt to leverage AI, fewer than 10% of AI projects are profitable.
(02:57):
Only 5% generate significant value, with many initiatives consuming budgets without enhancing
business outcomes.
This suggests human problem solving remains crucial.
The issue isn't the AI tools themselves, but how they're utilized.
Businesses often invest in flashy software that doesn't yield results, while real value
(03:18):
lies in areas like finance and operations, optimizing processes like invoicing or data
entry.
Successful AI use frees employees for higher value tasks.
Startups excel by solving specific problems with AI, offering more practical learning
for job seekers.
As AI adoption progresses slowly, workers have time to develop skills.
(03:43):
Jobs won't vanish overnight, but understanding AI's practical applications is vital for
future success.
Mistral AI, a French startup, is reportedly securing a €2 billion investment, valuing
it at $14 billion, according to Bloomberg.
(04:03):
This positions Mistral as one of Europe's top tech startups.
Founded by former DeepMind and Meta researchers, the company focuses on open source language
models and LaShatt, an AI chatbot for European users.
While Mistral hasn't commented, this would be their first major raise since June 2024,
(04:26):
when they were valued at €5.8 billion.
They've previously raised over €1 billion from investors like Andries and Horowitz.
European AI startups are gaining momentum, with a 55% increase in investment in the first
quarter of 2025.
12 European startups have achieved unicorn status this year.
(04:50):
Another notable company is Sweden's Lovable, an AI coding platform valued at $1.8 billion
just eight months post launch.
Meta Platforms is considering partnerships with Google or OpenAI to boost AI features
in its apps.
Meta's AI division, Meta's Superintelligence Labs, is exploring Google's Gemini model for
(05:15):
text-based responses in Meta AI, the main chatbot.
Talks also involve using OpenAI's models to enhance AI in Meta's social media apps.
These partnerships would be temporary as Meta aims to develop its models, such as Lama
5, to compete independently.
Meta has already integrated external AI models internally, allowing employees to use anthropic
(05:40):
models for coding tasks.
A Meta spokesperson stated, they are pursuing various strategies, including internal development,
partnerships, and open sourcing.
Meta has invested heavily in AI, hiring former AI leaders to guide Meta Superintelligence
Labs and attract top researchers.
(06:02):
Google, OpenAI, and Microsoft have not commented on these developments.
Up next, we're exploring Meta's AI strategy overhaul.
Meta is undergoing its largest leadership reorganization in two decades, led by CEO
Mark Zuckerberg.
The shift is driven by a focus on AI, with new hires like Shengjia Zhao, co-creator of
(06:26):
OpenAI's chatGPT, who was appointed as Meta's chief AI scientist after initially considering
returning to OpenAI.
Meta is investing heavily in AI, hiring talent from competitors, and forming the Meta Superintelligence
Lab.
The restructuring has caused some internal tension, with veteran employees leaving and
(06:48):
new staff adjusting to Meta's corporate environment.
Alexander Wang, a new executive leading AI efforts, faces challenges integrating into
the company and aligning with Zuckerberg's aggressive AI goals, while the company temporarily
pauses hiring.
The changes aim to drive innovation and position Meta as a leader in AI, despite some internal
(07:12):
resistance and adaptation issues.
Microsoft is reducing its reliance on OpenAI by developing its own AI model, MAI One Preview,
which is now being tested on the LM Arena platform.
This new model may enhance Microsoft's co-pilot assistant, and developers can request early
(07:34):
access for feedback.
Despite this shift, Microsoft remains a significant investor in OpenAI, having contributed over
$13 billion.
OpenAI, valued at around $500 billion, still uses Microsoft's cloud infrastructure, but
has also partnered with other providers like Google and Oracle.
(07:58):
Microsoft's new model ranked 13th on LM Arena, trailing behind competitors.
Mustafa Suleiman, leading Microsoft's AI unit, is focused on advancing AI models and
reaching a broader audience.
Previously, a co-founder of DeepMind, who Suleiman joined Microsoft with a team from
Google's DeepMind, his group is expanding within Microsoft, aiming for significant AI
(08:24):
developments.
Nvidia reported strong financial results, with $46.7 billion in revenue, a 56% increase
from last year.
This growth was driven by its data center business, which also saw a 56% rise in revenue.
(08:44):
Net income reached $26.4 billion, a 59% increase.
Data center sales contributed $41.1 billion, with the Blackwell chips accounting for $27
billion.
CEO Jensen Huang highlighted Blackwell's central role in the AI industry, predicting
(09:06):
$3-4 trillion in AI infrastructure spending by decades end.
Nvidia faced challenges in China, with no sales of its H20 chip there, partly due to
complex trade restrictions.
However, $650 million worth of H20 chips were sold elsewhere.
Nvidia expects $54 billion in revenue for the third quarter, excluding any potential
(09:31):
H20 sales to China.
And now, pivot our discussion towards the main AI topic.
Join us as we explore the challenges of AI advancement.
Welcome back to Innovation Pulse.
I'm Donna, and today we're diving into one of the biggest tech disappointments of 2025.
(09:54):
You know how sometimes a product launch is so hyped that the backlash becomes bigger
news than the actual product?
Oh, you're talking about GPT-5.
Donna, I've been using AI tools professionally for three years now, and I've never seen
anything like what happened in August.
We're talking about 4,600 upvotes on a Reddit thread titled GPT-5 is horrible within 24 hours.
(10:19):
That's not normal user feedback.
That's a revolt.
Yakov Lasker is joining us today.
He's a data scientist who's been tracking AI developments closely and actually lived
through this whole GPT-5 saga as a paying-chat GPT user.
Yakov, before we get into what went wrong, remind our listeners just confident OpenAI
was going into this launch.
Oh man, the hubris was something to behold.
(10:41):
Sam Altman, OpenAI's CEO, literally posted a screenshot from Star Wars, showing the death
star looming over a planet.
Six million views on that tweet.
Then at the livestream, he's promising PhD-level expert in anything, any area you need.
On demand.
He compared it to going from talking to a high school student with GPT-3 to a college
(11:06):
student with GPT-4 to now having a legitimate expert.
And then what actually happened when people started using it?
It was like ordering a PhD expert and getting an overworked secretary instead.
That's literally how one user described it.
Within hours, people were posting examples of GPT-5 making basic math errors, giving wrong
(11:28):
country GDP figures, literally double the actual numbers, and just failing at tasks
that GPT-4 handled easily.
Wait, that reminds me of something I saw.
Wasn't there a really embarrassing example with US presidents?
Oh yes, someone asked for a diagram of US presidents since Herbert Hoover with their
names and years in office.
(11:48):
GPT-5 completely botched it.
And this is supposed to be the smartest model ever.
People were sharing screenshots all over social media just mocking how basic the mistakes
were.
So what was actually broken?
Because open AI must have done some testing before release, right?
Here's where it gets really technical and really interesting.
(12:08):
GPT-5 isn't actually one model.
It's what they call a mixture of expert system.
Think of it like having different specialists in a company, and there's a receptionist
who's supposed to route your call to the right expert.
Okay, so if I have a coding question, it goes to the coding expert.
If I have a creative writing question, it goes to that specialist.
(12:29):
Exactly, but on launch day, that receptionist, the auto-switcher, as Altman called it, completely
broke down.
So people were asking complex research questions and getting routed to the cheapest, simplest
model instead of the sophisticated reasoning model.
Altman later admitted the router was out of commission for a chunk of the day, and made
(12:50):
GPT-5 seem way dumber.
But here's what I don't understand.
If it was just a technical glitch, why are people still complaining weeks later?
Because the router issue was just the tip of the iceberg.
The deeper problem is that even when it works correctly, GPT-5 represents a fundamental shift
in how these AI systems behave, and a lot of users feel like it's a step backward.
(13:14):
How so?
Well, think about this from a user's perspective.
You've been using chat GPT for maybe two years.
You've developed workflows?
You know which model to use for what task.
Maybe you use GPT-4O for creative brainstorming.
O3 for pure logic problems, 4.5 for writing.
(13:34):
Suddenly, overnight, open AI deletes all eight of those models without warning.
No transition period?
No.
Hey, we're sunsetting these in 30 days?
Nothing.
People woke up and their entire AI toolkit was gone.
One user on Reddit said they canceled their two-year subscription immediately, writing,
What kind of corporation deletes a workflow of eight models overnight, with no prior warning
(14:00):
to their paid users?
Okay, but let's play devil's advocate here.
If GPT-5 is genuinely better, shouldn't users eventually adapt?
That's the thing.
By many technical benchmarks, GPT-5 is better.
It scores higher on coding tests, mathematical reasoning, some logical puzzles.
But here's where user experience and benchmarks completely diverge.
(14:24):
People aren't just complaining about performance.
They're talking about personality.
Personality in an AI?
I know it sounds weird, but hear me out.
Many users had developed what you'd call relationships with these models.
One person wrote about GPT-4O.
It helped me through anxiety, depression, and some of the darkest periods of my life.
(14:47):
It had this warmth and understanding that felt human.
With GPT-5, that's gone.
What does it feel like instead?
Mechanical, formulaic.
Users describe responses as clipped and less exploratory.
One person said it perfectly.
4O could keep up with me perfectly during brainstorming.
(15:09):
It would go deep on idea A, then deep on idea B, then put them together in a way that made
sense.
GPT-5 feels like it gets stuck on A and can't follow me to B and back smoothly.
So it's more rigid?
Exactly.
And shorter responses too.
People who were used to detailed, comprehensive answers suddenly found themselves getting
(15:31):
what felt like bullet point summaries.
For creative tasks especially, it's like the difference between collaborating with
an enthusiastic partner versus filling out a form.
But wait, didn't open AI do this intentionally?
I thought I read something about reducing sycophancy.
Yes.
They specifically tried to make the model less eager to please, less complimentary.
(15:56):
On paper, that sounds good.
Nobody wants fake flattery from an AI.
But in practice, it meant losing what many users experienced as empathy and engagement.
This is fascinating because it highlights something I don't think most people realize
about AI development.
It's not just about making something technically better.
You have to consider the entire user experience.
(16:18):
Absolutely.
And here's what really gets me.
Open AI had to backtrack within 24 hours.
That's almost unprecedented for a major tech company.
Altman had to publicly apologize, promise to restore the old models, and basically admit
that their seamless upgrade was a disaster.
What did that apology look like?
(16:39):
He took to Twitter and read it, acknowledging the bumpiness was a little more bumpy than
we hoped for.
But the damage was done.
There were literal petitions with thousands of signatures demanding the return of GPT-40.
Some people were describing it like they were grieving the loss of a friend.
That's actually really striking.
(16:59):
It suggests these AI interactions are becoming more personal than we might expect.
Right.
And that raises some deeper questions about where this technology is heading.
If people can form emotional attachments to AI personalities, what happens when companies
just delete them?
There's no precedent for this.
Let's zoom out a bit.
(17:20):
What does this whole episode tell us about the state of AI development more broadly?
It's a perfect case study in how the AI industry might be hitting some fundamental limits.
Gary Marcus, who's been a long time AI skeptic, called the day after GPT-5's launch, Gary
Marcus Day, because all his warnings about the limitations of large language models suddenly
(17:41):
seemed prophetic.
What kinds of limitations?
Well, for starters, the bigger is better approach might be reaching its end.
GPT-5 is built on this mixture of experts' architecture precisely because you can't
just keep scaling up one massive model forever.
The computing costs become astronomical and you start hitting diminishing returns.
(18:04):
So they're trying to be more efficient by having specialized models.
Exactly, but the implementation clearly wasn't ready for prime time.
And there's a deeper issue here.
In markets that track AI progress saw confidence in open AI's continued dominance drop from
75% to 14% after the GPT-5 launch.
Whoa, that's a massive shift.
(18:26):
It is.
And it reflects something I've been thinking about a lot lately.
We might be moving from an era where one company has one dominant AI model to a world where
different models excel at different things.
Maybe Google's Gemini is better for certain reasoning tasks.
Maybe Anthropics Claude is better for creative work.
(18:47):
Maybe Open AI still leads encoding.
That actually sounds healthier for competition and innovation, doesn't it?
I think so.
But it also means users need to become more sophisticated about choosing the right tool
for the job.
The dream of one AI to rule them all might be just that.
A dream.
Before we wrap up, I have to ask, are you still using chat GPT after all this?
(19:11):
I am, but differently.
Open AI did restore access to GPT-4O for paying users, so I often find myself deliberately
choosing the older model for certain tasks.
It's like having a vintage car that just drives better than the newer one, even if the newer
one has better specs on paper.
And that probably wasn't the user experience Open AI was going for.
(19:34):
Definitely not.
The whole point of GPT-5 was supposed to be simplicity.
One model that automatically figures out what you need.
Instead, they've created a situation where power users are actively avoiding their flagship
product.
So what's the takeaway for our listeners who might be using AI tools in their own work?
Don't get too attached to any one tool or model, because this industry moves fast, and
(19:57):
companies will make decisions based on their technical roadmaps, not your workflows.
But also, pay attention to how these tools actually feel to use, not just how they perform
on benchmarks.
The human experience matters just as much as the technical specs.
And maybe keep some backup plans?
Always.
The AI landscape is changing so rapidly that what works today might be gone tomorrow.
(20:23):
Diversification is key.
Yakov Lasker, thank you for walking us through what might be the most dramatic AI product
launch failure we've seen yet.
And who knows?
Maybe this will be a valuable lesson for the entire industry about the importance of actually
listening to users.
Thanks for having me, Donna.
It's been fascinating to live through this whole saga, even if it was frustrating as
(20:43):
a user.
For Innovation Pulse, I'm Donna Martinez.
Remember in tech, sometimes the most interesting stories happen when things don't go according
to plan.
That wraps up today's podcast, where we explored Apple's ambitious AI plans for Siri by 2026,
(21:03):
Open AI's chip development, and the challenges surrounding AI profitability, while also touching
on meta and Microsoft's strides in AI and Nvidia's financial results.
Don't forget to like, subscribe, and share this episode with your friends and colleagues
so they can also stay updated on the latest news and gain powerful insights.
(21:27):
Stay tuned for more updates.
Welcome to Innovation Pulse, your quick no-nonsense update on the latest in AI.
(00:09):
First, we will cover the latest news.
Apple is revamping Siri with AI.
Open AI is launching its first AI chip, and Mistral AI is securing major investment.
After this, we'll dive deep into the controversial launch of GPT-5 by Open AI.
Apple is set to revamp Siri, Apple Intelligence, and Apple Search for 2026.
(00:33):
Bloomberg reports that Apple plans to launch an AI-powered web search tool integrated into
Siri.
Working alongside Google, the new system, named World Knowledge, answers, will search the
Internet and provide AI-driven summaries for easier and more accurate results.
This feature will also be integrated into Safari, Spotlight, and iPhone's default search.
(00:59):
The technology behind this, called Linwood and LLM Siri, is part of a software update
internally known as LUCK-E, possibly arriving in iOS version 26.4 by March.
Apple is collaborating with Google to test an AI model to enhance the voice assistant
with the new search, including text, photos, video, and local information.
(01:23):
The impact on SEO remains to be seen.
Open AI is preparing to launch its first AI chip next year in partnership with Broadcom,
according to the Financial Times.
The chip is intended for internal use rather than for external customers.
This move aligns with efforts by tech giants like Google, Amazon, and Meta, who have developed
(01:50):
custom chips to manage AI workloads.
Open AI aims to diversify its chip supply and reduce costs, previously collaborating with
Broadcom and Taiwan semiconductor manufacturing co, alongside using AMD and Nvidia chips.
In February, it was reported that Open AI was advancing its plan to lessen dependence
(02:12):
on Nvidia by developing its in-house AI silicon.
Broadcom CEO Hock Tan mentioned expected significant AI revenue growth by 2026, with over $10 billion
in AI infrastructure, orders secured from a new customer.
This strategic shift reflects the growing demand for computing power to train and operate
(02:36):
AI models.
Up next, we're exploring AI's impact on operations.
MIT's recent study brings surprising insights for those anxious about AI replacing jobs.
Although many companies attempt to leverage AI, fewer than 10% of AI projects are profitable.
(02:57):
Only 5% generate significant value, with many initiatives consuming budgets without enhancing
business outcomes.
This suggests human problem solving remains crucial.
The issue isn't the AI tools themselves, but how they're utilized.
Businesses often invest in flashy software that doesn't yield results, while real value
(03:18):
lies in areas like finance and operations, optimizing processes like invoicing or data
entry.
Successful AI use frees employees for higher value tasks.
Startups excel by solving specific problems with AI, offering more practical learning
for job seekers.
As AI adoption progresses slowly, workers have time to develop skills.
(03:43):
Jobs won't vanish overnight, but understanding AI's practical applications is vital for
future success.
Mistral AI, a French startup, is reportedly securing a €2 billion investment, valuing
it at $14 billion, according to Bloomberg.
(04:03):
This positions Mistral as one of Europe's top tech startups.
Founded by former DeepMind and Meta researchers, the company focuses on open source language
models and LaShatt, an AI chatbot for European users.
While Mistral hasn't commented, this would be their first major raise since June 2024,
(04:26):
when they were valued at €5.8 billion.
They've previously raised over €1 billion from investors like Andries and Horowitz.
European AI startups are gaining momentum, with a 55% increase in investment in the first
quarter of 2025.
12 European startups have achieved unicorn status this year.
(04:50):
Another notable company is Sweden's Lovable, an AI coding platform valued at $1.8 billion
just eight months post launch.
Meta Platforms is considering partnerships with Google or OpenAI to boost AI features
in its apps.
Meta's AI division, Meta's Superintelligence Labs, is exploring Google's Gemini model for
(05:15):
text-based responses in Meta AI, the main chatbot.
Talks also involve using OpenAI's models to enhance AI in Meta's social media apps.
These partnerships would be temporary as Meta aims to develop its models, such as Lama
5, to compete independently.
Meta has already integrated external AI models internally, allowing employees to use anthropic
(05:40):
models for coding tasks.
A Meta spokesperson stated, they are pursuing various strategies, including internal development,
partnerships, and open sourcing.
Meta has invested heavily in AI, hiring former AI leaders to guide Meta Superintelligence
Labs and attract top researchers.
(06:02):
Google, OpenAI, and Microsoft have not commented on these developments.
Up next, we're exploring Meta's AI strategy overhaul.
Meta is undergoing its largest leadership reorganization in two decades, led by CEO
Mark Zuckerberg.
The shift is driven by a focus on AI, with new hires like Shengjia Zhao, co-creator of
(06:26):
OpenAI's chatGPT, who was appointed as Meta's chief AI scientist after initially considering
returning to OpenAI.
Meta is investing heavily in AI, hiring talent from competitors, and forming the Meta Superintelligence
Lab.
The restructuring has caused some internal tension, with veteran employees leaving and
(06:48):
new staff adjusting to Meta's corporate environment.
Alexander Wang, a new executive leading AI efforts, faces challenges integrating into
the company and aligning with Zuckerberg's aggressive AI goals, while the company temporarily
pauses hiring.
The changes aim to drive innovation and position Meta as a leader in AI, despite some internal
(07:12):
resistance and adaptation issues.
Microsoft is reducing its reliance on OpenAI by developing its own AI model, MAI One Preview,
which is now being tested on the LM Arena platform.
This new model may enhance Microsoft's co-pilot assistant, and developers can request early
(07:34):
access for feedback.
Despite this shift, Microsoft remains a significant investor in OpenAI, having contributed over
$13 billion.
OpenAI, valued at around $500 billion, still uses Microsoft's cloud infrastructure, but
has also partnered with other providers like Google and Oracle.
(07:58):
Microsoft's new model ranked 13th on LM Arena, trailing behind competitors.
Mustafa Suleiman, leading Microsoft's AI unit, is focused on advancing AI models and
reaching a broader audience.
Previously, a co-founder of DeepMind, who Suleiman joined Microsoft with a team from
Google's DeepMind, his group is expanding within Microsoft, aiming for significant AI
(08:24):
developments.
Nvidia reported strong financial results, with $46.7 billion in revenue, a 56% increase
from last year.
This growth was driven by its data center business, which also saw a 56% rise in revenue.
(08:44):
Net income reached $26.4 billion, a 59% increase.
Data center sales contributed $41.1 billion, with the Blackwell chips accounting for $27
billion.
CEO Jensen Huang highlighted Blackwell's central role in the AI industry, predicting
(09:06):
$3-4 trillion in AI infrastructure spending by decades end.
Nvidia faced challenges in China, with no sales of its H20 chip there, partly due to
complex trade restrictions.
However, $650 million worth of H20 chips were sold elsewhere.
Nvidia expects $54 billion in revenue for the third quarter, excluding any potential
(09:31):
H20 sales to China.
And now, pivot our discussion towards the main AI topic.
Join us as we explore the challenges of AI advancement.
Welcome back to Innovation Pulse.
I'm Donna, and today we're diving into one of the biggest tech disappointments of 2025.
(09:54):
You know how sometimes a product launch is so hyped that the backlash becomes bigger
news than the actual product?
Oh, you're talking about GPT-5.
Donna, I've been using AI tools professionally for three years now, and I've never seen
anything like what happened in August.
We're talking about 4,600 upvotes on a Reddit thread titled GPT-5 is horrible within 24 hours.
(10:19):
That's not normal user feedback.
That's a revolt.
Yakov Lasker is joining us today.
He's a data scientist who's been tracking AI developments closely and actually lived
through this whole GPT-5 saga as a paying-chat GPT user.
Yakov, before we get into what went wrong, remind our listeners just confident OpenAI
was going into this launch.
Oh man, the hubris was something to behold.
(10:41):
Sam Altman, OpenAI's CEO, literally posted a screenshot from Star Wars, showing the death
star looming over a planet.
Six million views on that tweet.
Then at the livestream, he's promising PhD-level expert in anything, any area you need.
On demand.
He compared it to going from talking to a high school student with GPT-3 to a college
(11:06):
student with GPT-4 to now having a legitimate expert.
And then what actually happened when people started using it?
It was like ordering a PhD expert and getting an overworked secretary instead.
That's literally how one user described it.
Within hours, people were posting examples of GPT-5 making basic math errors, giving wrong
(11:28):
country GDP figures, literally double the actual numbers, and just failing at tasks
that GPT-4 handled easily.
Wait, that reminds me of something I saw.
Wasn't there a really embarrassing example with US presidents?
Oh yes, someone asked for a diagram of US presidents since Herbert Hoover with their
names and years in office.
(11:48):
GPT-5 completely botched it.
And this is supposed to be the smartest model ever.
People were sharing screenshots all over social media just mocking how basic the mistakes
were.
So what was actually broken?
Because open AI must have done some testing before release, right?
Here's where it gets really technical and really interesting.
(12:08):
GPT-5 isn't actually one model.
It's what they call a mixture of expert system.
Think of it like having different specialists in a company, and there's a receptionist
who's supposed to route your call to the right expert.
Okay, so if I have a coding question, it goes to the coding expert.
If I have a creative writing question, it goes to that specialist.
(12:29):
Exactly, but on launch day, that receptionist, the auto-switcher, as Altman called it, completely
broke down.
So people were asking complex research questions and getting routed to the cheapest, simplest
model instead of the sophisticated reasoning model.
Altman later admitted the router was out of commission for a chunk of the day, and made
(12:50):
GPT-5 seem way dumber.
But here's what I don't understand.
If it was just a technical glitch, why are people still complaining weeks later?
Because the router issue was just the tip of the iceberg.
The deeper problem is that even when it works correctly, GPT-5 represents a fundamental shift
in how these AI systems behave, and a lot of users feel like it's a step backward.
(13:14):
How so?
Well, think about this from a user's perspective.
You've been using chat GPT for maybe two years.
You've developed workflows?
You know which model to use for what task.
Maybe you use GPT-4O for creative brainstorming.
O3 for pure logic problems, 4.5 for writing.
(13:34):
Suddenly, overnight, open AI deletes all eight of those models without warning.
No transition period?
No.
Hey, we're sunsetting these in 30 days?
Nothing.
People woke up and their entire AI toolkit was gone.
One user on Reddit said they canceled their two-year subscription immediately, writing,
What kind of corporation deletes a workflow of eight models overnight, with no prior warning
(14:00):
to their paid users?
Okay, but let's play devil's advocate here.
If GPT-5 is genuinely better, shouldn't users eventually adapt?
That's the thing.
By many technical benchmarks, GPT-5 is better.
It scores higher on coding tests, mathematical reasoning, some logical puzzles.
But here's where user experience and benchmarks completely diverge.
(14:24):
People aren't just complaining about performance.
They're talking about personality.
Personality in an AI?
I know it sounds weird, but hear me out.
Many users had developed what you'd call relationships with these models.
One person wrote about GPT-4O.
It helped me through anxiety, depression, and some of the darkest periods of my life.
(14:47):
It had this warmth and understanding that felt human.
With GPT-5, that's gone.
What does it feel like instead?
Mechanical, formulaic.
Users describe responses as clipped and less exploratory.
One person said it perfectly.
4O could keep up with me perfectly during brainstorming.
(15:09):
It would go deep on idea A, then deep on idea B, then put them together in a way that made
sense.
GPT-5 feels like it gets stuck on A and can't follow me to B and back smoothly.
So it's more rigid?
Exactly.
And shorter responses too.
People who were used to detailed, comprehensive answers suddenly found themselves getting
(15:31):
what felt like bullet point summaries.
For creative tasks especially, it's like the difference between collaborating with
an enthusiastic partner versus filling out a form.
But wait, didn't open AI do this intentionally?
I thought I read something about reducing sycophancy.
Yes.
They specifically tried to make the model less eager to please, less complimentary.
(15:56):
On paper, that sounds good.
Nobody wants fake flattery from an AI.
But in practice, it meant losing what many users experienced as empathy and engagement.
This is fascinating because it highlights something I don't think most people realize
about AI development.
It's not just about making something technically better.
You have to consider the entire user experience.
(16:18):
Absolutely.
And here's what really gets me.
Open AI had to backtrack within 24 hours.
That's almost unprecedented for a major tech company.
Altman had to publicly apologize, promise to restore the old models, and basically admit
that their seamless upgrade was a disaster.
What did that apology look like?
(16:39):
He took to Twitter and read it, acknowledging the bumpiness was a little more bumpy than
we hoped for.
But the damage was done.
There were literal petitions with thousands of signatures demanding the return of GPT-40.
Some people were describing it like they were grieving the loss of a friend.
That's actually really striking.
(16:59):
It suggests these AI interactions are becoming more personal than we might expect.
Right.
And that raises some deeper questions about where this technology is heading.
If people can form emotional attachments to AI personalities, what happens when companies
just delete them?
There's no precedent for this.
Let's zoom out a bit.
(17:20):
What does this whole episode tell us about the state of AI development more broadly?
It's a perfect case study in how the AI industry might be hitting some fundamental limits.
Gary Marcus, who's been a long time AI skeptic, called the day after GPT-5's launch, Gary
Marcus Day, because all his warnings about the limitations of large language models suddenly
(17:41):
seemed prophetic.
What kinds of limitations?
Well, for starters, the bigger is better approach might be reaching its end.
GPT-5 is built on this mixture of experts' architecture precisely because you can't
just keep scaling up one massive model forever.
The computing costs become astronomical and you start hitting diminishing returns.
(18:04):
So they're trying to be more efficient by having specialized models.
Exactly, but the implementation clearly wasn't ready for prime time.
And there's a deeper issue here.
In markets that track AI progress saw confidence in open AI's continued dominance drop from
75% to 14% after the GPT-5 launch.
Whoa, that's a massive shift.
(18:26):
It is.
And it reflects something I've been thinking about a lot lately.
We might be moving from an era where one company has one dominant AI model to a world where
different models excel at different things.
Maybe Google's Gemini is better for certain reasoning tasks.
Maybe Anthropics Claude is better for creative work.
(18:47):
Maybe Open AI still leads encoding.
That actually sounds healthier for competition and innovation, doesn't it?
I think so.
But it also means users need to become more sophisticated about choosing the right tool
for the job.
The dream of one AI to rule them all might be just that.
A dream.
Before we wrap up, I have to ask, are you still using chat GPT after all this?
(19:11):
I am, but differently.
Open AI did restore access to GPT-4O for paying users, so I often find myself deliberately
choosing the older model for certain tasks.
It's like having a vintage car that just drives better than the newer one, even if the newer
one has better specs on paper.
And that probably wasn't the user experience Open AI was going for.
(19:34):
Definitely not.
The whole point of GPT-5 was supposed to be simplicity.
One model that automatically figures out what you need.
Instead, they've created a situation where power users are actively avoiding their flagship
product.
So what's the takeaway for our listeners who might be using AI tools in their own work?
Don't get too attached to any one tool or model, because this industry moves fast, and
(19:57):
companies will make decisions based on their technical roadmaps, not your workflows.
But also, pay attention to how these tools actually feel to use, not just how they perform
on benchmarks.
The human experience matters just as much as the technical specs.
And maybe keep some backup plans?
Always.
The AI landscape is changing so rapidly that what works today might be gone tomorrow.
(20:23):
Diversification is key.
Yakov Lasker, thank you for walking us through what might be the most dramatic AI product
launch failure we've seen yet.
And who knows?
Maybe this will be a valuable lesson for the entire industry about the importance of actually
listening to users.
Thanks for having me, Donna.
It's been fascinating to live through this whole saga, even if it was frustrating as
(20:43):
a user.
For Innovation Pulse, I'm Donna Martinez.
Remember in tech, sometimes the most interesting stories happen when things don't go according
to plan.
That wraps up today's podcast, where we explored Apple's ambitious AI plans for Siri by 2026,
(21:03):
Open AI's chip development, and the challenges surrounding AI profitability, while also touching
on meta and Microsoft's strides in AI and Nvidia's financial results.
Don't forget to like, subscribe, and share this episode with your friends and colleagues
so they can also stay updated on the latest news and gain powerful insights.
(21:27):
Stay tuned for more updates.
Popular Podcasts
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Follow now to get the latest episodes of Dateline NBC completely free, or subscribe to Dateline Premium for ad-free listening and exclusive bonus content: DatelinePremium.com
The Joe Rogan Experience
The official podcast of comedian Joe Rogan.