15 August 2023

Tech Expert Reveals How AI Could Destroy Your Startup (#147)

2Frameworks

12Insights

Frameworks in this episode

Strategy6 steps

Generalist-to-Specialist Model Routing

Route each AI task to the smallest model suited to its domain

Read →

Innovation5 steps

Lowest-Hanging-Fruit Learning Build

Build the simplest useful product to learn an unfamiliar technology fast

Read →

Insights & moments

The myth-busts, hot takes, explainers, and tools worth keeping.

Myth Buster· 1

Myth Buster13:00

Why an 80% Convincing AI Caller Is Not Production Ready

The hosts use AI phone agents to illustrate the difference between an impressive demonstration and a dependable business system. Reaching roughly human-like speech is not enough if latency, problem resolution, or edge-case behavior still makes the interaction unsuitable for serious sales and support work.

A convincing demo can remain unusable in customer-facing operations
The final quality gap may be harder than the initial breakthrough
Latency makes synthetic conversations feel unnatural
Better underlying models can improve a product without changing its interface
Local infrastructure may reduce response delays

“that 80% is cool, but not something any serious company would use in their business.”
Kieran Flanagan · 13:30

“people underestimate how hard it is to make the last bit, like the last 10% to become magical.”
Linus Ekenstam · 14:30

#voice ai#customer support#sales#latency

Hot Take· 5

Hot Take07:30

The Real AI Magic Is Below the Chat Interface

Ekenstam distinguishes consumer chat products from the deeper application potential of model APIs. APIs can code, generate, compress, transform, and translate many items concurrently, allowing builders to compose capabilities into products that go far beyond a one-to-one chat exchange.

Chat interfaces reveal only part of model capability
APIs support transformation, translation, compression, generation, and coding
Multiple AI operations can run in parallel
Application-level composition offers more durable opportunity than consumer novelty

“you can just send something in and it can code for you, it can generate, it can compress, it can transform, it can translate, and…”
Linus Ekenstam · 08:00

“the stuff you can build with these tools, that is where the magic is.”
Linus Ekenstam · 08:30

#apis#ai applications#automation#product development

Hot Take17:00

AI Could Shift the Economy Toward Smaller Companies

Kipp argues that AI's largest economic effect may not be incremental productivity inside large corporations. Instead, automation and synthetic labor could let very small teams create viable businesses in markets where the previous staffing economics made entry impossible.

AI may replace substantial portions of knowledge work
Small teams could operate at previously impossible productivity levels
New business formation may matter more than efficiency gains at incumbents
AI-mediated interaction could change how companies earn revenue

“the future of our economy is more smaller businesses”
Kipp Bodner · 17:30

“AI is going to enable you to build a really productive business in areas that weren't possible before with very few people.”
Kipp Bodner · 17:30

#small business#entrepreneurship#future of work#economics

Hot Take30:30

Four Visitors Could See Four Different Versions of One Website

Ekenstam predicts that websites will remain during a gradual transition but become more dynamic, automatically maintained, and personalized. Copy and images may initially change within a stable layout, while richer adaptation could eventually respond to each visitor's data, needs, and context.

Spatial computing may change the screen-based web gradually
Familiar windows will persist during the transition
Agents could help build and maintain sites continuously
Copy and imagery are immediate personalization opportunities
Different visitors may receive substantially different experiences

“website will be around, but they will radically change how they're made, how they're maintained”
Linus Ekenstam · 32:00

“imagine the four of us browsing the same website, but seeing four very different websites.”
Linus Ekenstam · 32:30

#websites#personalization#spatial computing#web design

Hot Take33:00

A Universal Copilot May Matter More Than a Chatbot on Every Site

Ekenstam argues that speaking separately to an unfamiliar, forgetful AI on every website would become tiring. A system-level copilot that knows the user and can interact with any website could instead connect personal preferences to site-specific services and interfaces.

A separate chat experience on every site creates friction
Site-specific assistants may lack memory of the user
An operating-system copilot could work across applications and websites
Websites could expose capabilities to a trusted personal assistant
Copilot adoption may reshape marketing collateral and interfaces

“If everything moves into chat, that's a pretty boring future.”
Linus Ekenstam · 33:30

“I do believe harder in the notion of like having a copilot as in the operating system”
Linus Ekenstam · 33:30

#copilots#websites#personalization#operating systems

Hot Take35:00

Routine Purchases May Stop Needing a Visual Website

Ekenstam expects interfaces to fade for routine purchases where consumers do not need to inspect the product. Assistants could reorder commodities or communicate with utility providers directly, while visual interfaces remain important for products such as fashion where appearance drives the decision.

Voice assistants already reduce the need to browse for routine goods
Commodity purchases can happen without viewing a product page
Visual interfaces remain valuable when appearance affects choice
Audio delivery can replace websites for some information consumption
The web's medium may change more than it has since its creation

“I think technology UI will slowly fade away.”
Linus Ekenstam · 35:00

“I never went anywhere to buy that toilet paper, right?”
Linus Ekenstam · 35:30

#commerce#voice assistants#user interfaces#websites

Explainer· 3

Explainer05:00

Why ChatGPT's Interface Made Existing AI Feel Revolutionary

Kieran argues that ChatGPT's breakthrough was not solely a new underlying API but an interface that made conversational AI accessible to a broad audience. Its fast feedback and back-and-forth interaction could feel more natural than retrieving information through conventional search links.

Accessible UX exposed existing capabilities to far more people
Conversation enabled rapid learning and iteration
Chat-based retrieval sometimes felt better than traditional search
The initial adoption curve encouraged unrealistic expectations about the pace of change

“the UX had made it much more easier for like broad study people to use.”
Kieran Flanagan · 05:00

“it's the speed of learning and iteration.”
Kieran Flanagan · 05:00

#chatgpt#user experience#search#ai adoption

Explainer10:00

Multimodal AI Gives a Text Model Eyes on the Physical World

Ekenstam describes a text-only chat model as effectively blind to the user's surroundings. Image input lets a model interpret visible objects and documents, while image output and native audio expand the number of tasks it can perform across cooking, software creation, and health-related assistance.

Text-only interaction limits models to what users describe
Image input connects AI to visible objects and documents
Image generation expands possible outputs
Native image and audio processing can remove intermediate conversion steps
Health uses require professional confirmation rather than unquestioned model trust

“it's like you are talking to a blind person.”
Linus Ekenstam · 10:00

“the moment it can take an image as an input, it's like giving the model ice.”
Linus Ekenstam · 10:30

#multimodal ai#computer vision#audio#healthcare

Explainer18:30

Why GPT-4 Could Seem Worse Without a Simple Intelligence Decline

The episode examines reported performance drops in ChatGPT and several possible causes, including reduced context length, cost optimization, model routing, and a product design that expects more structured instructions. The discussion emphasizes that ChatGPT's opaque consumer interface does not reveal which model or context configuration produced an answer.

Reported benchmark accuracy fell sharply on selected tasks
A shorter context window can make long chats appear forgetful
Opaque routing may send a task to a less capable model
Changes intended to reduce computational cost may affect quality
Consumer ChatGPT behavior should not be conflated with every API model

“responses that were directly executable decreased from 52% in March to 10% in June.”
Kieran Flanagan · 19:00

“it's very non-transparent what model runs.”
Linus Ekenstam · 21:30

#gpt-4#model quality#context windows#benchmarks

Story· 1

Story02:30

The Three Months That Turned AI Curiosity Into Commitment

Ekenstam says his transition to going all in on AI occurred across September, October, and November rather than in one dramatic moment. Repeated experiences of completing weeks of work in hours, learning through argument, and building functioning concepts convinced him that AI would affect nearly every kind of job.

The tipping point accumulated across several months
AI compressed some tasks from weeks into hours
Conversational disagreement produced new perspectives and better outputs
Sharing practical lessons publicly reinforced his commitment

“It wasn't a singular event.”
Linus Ekenstam · 03:00

“The genie's out of the bottle, I'm just gonna ride this wave.”
Linus Ekenstam · 04:30

#ai adoption#career change#learning#productivity

Takeaway· 2

Takeaway19:30

AI Wrappers Can Improve or Break Overnight

Companies built primarily around third-party models do not fully control their product experience. An upstream model upgrade may improve them immediately, but degradation, routing changes, or platform competition can also weaken their output without a corresponding change in their own code.

Third-party models can change product behavior overnight
Low barriers to entry make wrapper products easy to copy
Model providers can absorb wrapper features into their platforms
Incumbents have responded to AI competition by shipping faster
Dependence is dangerous when upstream behavior is not observable

“their entire product will be updated overnight and get better overnight.”
Kieran Flanagan · 19:30

“But the same is also true. Like it can get much worse.”
Kieran Flanagan · 19:30

#startups#platform risk#ai wrappers#competition

Takeaway29:00

Businesses Trust AI Internally Before Exposing It to Customers

Zapier research suggested that customers remained hesitant to put uncertain AI systems into customer-facing interactions. Companies were more comfortable using AI internally for writing, video production, research, and sales preparation, where employees can inspect the output before it affects a customer.

Customer-facing AI creates greater perceived business risk
Internal workflows allow human review before action
Writing and video acceleration are common early uses
Sales research can combine internal data into faster account briefs
Trust remains a barrier to business-critical deployment

“customers first are still hesitant to use AI for like customer facing experiences”
Kieran Flanagan · 29:00

“people are using it much more for like internal things where they feel much more safer”
Kieran Flanagan · 29:30

#enterprise ai#internal automation#customer experience#trust