PimEyes can locate photographs of someone across the internet from a single uploaded image. The hosts highlight both its remarkable search capability and the troubling requirement to submit identification when requesting removal.

One photograph can surface older and unexpected images of a person
Facial search demonstrates how search is moving beyond text
The opt-out process reportedly requires an ID or passport
Privacy remedies can introduce additional privacy risks

“You upload one picture of somebody and it finds all the pictures of them on the entire internet.”
Kip Bodner · 02:00

“you can opt out of being included in this search, but you must upload your ID or passport to verify.”
Kip Bodner · 03:00

#facial-search#privacy#ai-search#pimeyes

Hot Take· 2

Hot Take10:00

Real-Time Translation Could Accelerate Globalization

Kip predicts that translation tools will deepen global economic participation by helping businesses in emerging markets reach English-speaking customers. The same technology could also make cross-language travel conversations substantially more natural than today's phone-based translation exchanges.

Translation can connect non-English-speaking businesses to larger markets
Emerging-market companies may gain easier access to English-speaking customers
Real-time speech translation could transform travel
Communication quality must improve beyond awkward turn-based phone translation

“In the next five years, we are going to become more of a global economy than we've ever been because of tools like this.”
Kip Bodner · 10:00

“not just better for the world. It's better that we can all communicate with each other, talk to each other, learn about each other, and…”
Kieran Flanagan · 12:30

#globalization#emerging-markets#travel#translation

Hot Take17:30

AI Is Collapsing the Creative Talent Stack

Kieran and Kip argue that generative tools let one person execute a broader creative vision without passing it through multiple specialists. They expect this reduction in production dependencies to produce more direct, distinctive, and individually authored work.

One creator can perform tasks previously divided among specialists
Fewer handoffs can preserve the originator's creative intent
Natural-language tools reduce technical production barriers
Individual creative output may become sharper and more distinctive

“We are a collapsing the talent stack.”
Kieran Flanagan · 17:30

“we are embarking on an era of the renaissance of individual creativity.”
Kip Bodner · 17:30

#creativity#future-of-work#marketing#generative-ai

Explainer· 4

Explainer03:00

Why Search Is Splintering Into Specialized AI Tools

The hosts frame facial recognition as an early example of search becoming multimodal and distributed across specialized applications. Future tools may search through images and videos rather than relying primarily on text queries.

Search is moving into specialized point-solution applications
Images can function as queries rather than merely results
Video-based discovery may become another major search modality
New search capabilities will support both beneficial and harmful uses

“Like that is a form of search, right? Like we're seeing search continue to be de aggregated, decommoditized, like live inside all of these singular…”
Kieran Flanagan · 03:00

“multimodal search is gonna be pretty fascinating to watch as it kind of unravels.”
Kieran Flanagan · 04:00

#multimodal-search#vertical-search#video-search#ai

Explainer06:30

Meta’s Seamless M4T Lowers the Cost of Going Global

Meta's Seamless M4T accepts speech or text across many languages and translates into a smaller supported set of output languages. The hosts argue that this can dramatically reduce the cost and delay involved in taking software and commerce businesses into international markets.

The model handles speech and text across roughly 100 input languages
Its output-language coverage is more limited than its input coverage
Translation can make marketing and support content globally accessible
Software and e-commerce businesses stand to benefit most immediately

“Meta released a brand new model called Seamless M4T.”
Kip Bodner · 06:30

“It can understand more languages than it can output, is basically one of the limitations of this model right now.”
Kip Bodner · 09:30

#meta#translation#globalization#seamless-m4t

Explainer18:00

Reverse Image Search May Help Expose Synthetic Identities

The hosts propose using image-search tools to assess whether an online persona is real or artificially generated. An image appearing only in one suspicious context may indicate a synthetic identity, although they caution that genuine people with little online presence can produce the same signal.

Image provenance can help evaluate suspicious online accounts
A one-of-one image may indicate an AI-generated persona
Fake advertisements and social accounts are potential applications
A limited online footprint is not conclusive proof of fabrication

“One of the biggest PIM eyes use cases is going to be determining if somebody's real or AI, right?”
Kip Bodner · 18:00

“this person is a one-of-one, more than likely fake versus like real.”
Kieran Flanagan · 19:00

#synthetic-media#identity#misinformation#image-search

Explainer19:30

Fine-Tuned GPT-3.5 Can Rival a Larger Generic Model

OpenAI's GPT-3.5 fine-tuning update allows businesses to adapt a general model to specialized data, examples, and operating requirements. The hosts say this can produce GPT-4-level performance in focused use cases while giving companies outputs better aligned with their products and brands.

Fine-tuning adapts a horizontal model to a company's specific use case
Business examples and proprietary operating patterns improve relevance
Specialized GPT-3.5 performance may rival generic GPT-4 performance
Brand voice, coding conventions, and support knowledge are strong applications

“GPT 3.5 performs as well as GPT 4 with this super fine tuning.”
Kip Bodner · 19:30

“tomorrow, you can actually fine-tune the LLM on your content, on your tone of voice, on your imagery, on your videos, and it's going to…”
Kieran Flanagan · 21:30

#openai#fine-tuning#gpt-3-5#enterprise-ai

Story· 2

Story05:00

How a Winning Ad Trapped HubSpot Into Reusing One Person’s Face

Kieran recounts how a high-performing HubSpot advertisement became difficult to replace after the advertising model learned its performance history. The featured employee consequently appeared across sales-tool advertising for years, illustrating the personal consequences of creative optimization.

A successful advertisement accumulated valuable performance history
Replacement creatives struggled to outperform the trained incumbent
Optimization incentives encouraged continued use of the same face
Marketers should consider the long-term consequences of featuring employees

“the conversion rate was really strong because the model had trained on that ad.”
Kieran Flanagan · 05:30

“We'd be conscious of the photos you're using in your ads.”
Kieran Flanagan · 06:00

#advertising#creative-testing#privacy#hubspot

Story22:30

HubSpot Saw a 10x Support-Answer Gain From Fine-Tuning

Kip describes HubSpot's experience tuning a model with detailed product data, examples, and customer-support use cases. He reports that this specialization substantially improved the quality of answers given to customers struggling with the software.

HubSpot built a model to respond to customer product questions
The model received granular product data and examples
Fine-tuning improved the relevance and quality of answers
Customer support is a high-value enterprise use case for specialized models

“fine tuning and really giving that model highly granular data and examples and use cases of those products, really increase the performance and the quality…”
Kip Bodner · 23:00

“Like it's incredible.”
Kip Bodner · 23:00

#hubspot#customer-support#fine-tuning#enterprise-ai

Tool· 1

Tool13:30

Midjourney Paint Turns Text Prompts Into Image Edits

Midjourney Paint allows users to select a region of a generated image and describe a replacement or modification in natural language. This addresses a major limitation of AI image generation by enabling rapid iteration without requiring a separate Photoshop workflow.

Users select a specific region of an image to modify
Natural-language instructions determine the requested change
Creators can iterate without exporting into traditional editing software
The feature supports fast visual experimentation and variation

“you basically can select a region of the image, then you can actually freeform type how you want that image to change.”
Kieran Flanagan · 14:30

“And now you can iterate through an image really quickly.”
Kip Bodner · 15:30

#midjourney#image-editing#generative-ai#design

Takeaway· 3

Takeaway07:30

Product-Led Companies Can Be International From Day One

Kieran argues that product-led businesses should translate customer-facing materials immediately rather than waiting for a traditional country-by-country expansion. Because freemium and touchless products already attract international users, localized documentation and support can help convert demand in non-English-speaking markets.

Begin with marketing collateral, documentation, and customer support
Use activation and upgrade rates to identify valuable localization opportunities
Freemium products naturally attract users outside their home market
Translation can accelerate entry into non-English-speaking countries

“You should be international from day one because you can actually have all of the material you need for your customers now in those languages…”
Kieran Flanagan · 08:00

“I think this is a way for you to actually lead in much faster to those non-English speaking countries if you're a PLG business.”
Kieran Flanagan · 09:00

#product-led-growth#localization#international-expansion#saas

Takeaway15:30

AI Image Editing Unlocks High-Volume Creative Testing

The hosts explain how rapid image editing can help advertisers produce and test numerous variations of the same concept. Teams can change people, clothing, colors, or other visual elements in minutes, reducing creative bottlenecks and replacing repetitive stock imagery with custom assets.

Advertising teams can generate hundreds of controlled variations
Small visual elements can be isolated for A/B testing
Paid teams no longer need to wait as long for creative production
Blogs can replace generic stock photography with unique images

“You can do that so quickly now, and you can do it so easily.”
Kip Bodner · 16:00

“Now they can do that, they can just rapidly have the AI create multiple versions of something and they can iterate and test something.”
Kieran Flanagan · 17:00

#creative-testing#paid-media#ab-testing#content-production

Takeaway24:30

Do Not Automatically Choose the Newest AI Model

The hosts recommend testing a fine-tuned smaller model against the latest general-purpose model rather than assuming newer means better. Context capacity, customization, privacy, security, cost, and task-specific outcomes can matter more than the model's generation number.

Benchmark multiple models on the actual business task
Compare fine-tuned GPT-3.5 with generic GPT-4
Evaluate outputs rather than relying on model prestige
Consider token limits and fine-tuning alongside raw model capability

“you might want to test 3.5 with super fine tuning. You might get a better output, which is very antithetical to how everybody thinks.”
Kip Bodner · 24:30

“So just don't opt for the newest model. Try a couple of different models, especially if you're an open AI user, you're using the API.”
Kip Bodner · 25:00

#model-selection#experimentation#openai-api#fine-tuning

Frameworks in this episode

Business-Specific LLM Fine-Tuning

Enterprise AI Adoption Gates

Natural-Language Creative Variant Testing

Outcome-Based LLM Selection

PLG International from Day One

Signal-Led International Expansion

Insights & moments

⚡Myth Buster· 1

PimEyes Makes Facial Search Powerful—and Unsettling

◆Hot Take· 2