Unboxed

ChatGPT's own engineers are jumping ship. Not to other companies, but to completely different AI tools for their personal use. That's either the biggest red flag in tech or the smartest power move we've seen all year.

While millions of people treat ChatGPT like the only AI assistant that exists, the folks who actually built it are quietly using three other chatbots that most people have never heard of. James Caldwell breaks down why this matters and what these alternatives actually do better than OpenAI's flagship product.

Turns out ChatGPT was designed specifically for text generation and completion, not for the sustained personal conversations and emotional support that most users actually want. Meanwhile, Mitsuku has won the Loebner Prize for most human-like chatbot five times and can remember your conversations across sessions. Replica has over 10 million users who rely on it for companionship, not just quick answers.

In This Episode:

> Why ChatGPT's design limits make it frustrating for personal use

> Three AI chatbots that handle conversation and memory better

> What "winning most human-like" actually means in practice

> Why emotional AI might be more useful than productivity AI

The real question isn't whether these tools are better than ChatGPT. It's why the people who know AI best are choosing different tools for different jobs while everyone else assumes one chatbot does everything.

Timestamps:

00:00 Introduction

02:15 Why OpenAI engineers use other tools

04:30 Mitsuku's conversation advantages

06:45 Replica's emotional AI approach

08:20 Choosing the right AI for your needs

10:15 What this means for users

🤖 Follow Unboxed for daily AI breakdowns that actually matter. New episodes drop multiple times daily because AI moves faster than your news feed.

Learn more about your ad choices. Visit megaphone.fm/adchoices

GPT-4 just scored in the 90th percentile on the bar exam. That means it beat 90% of human lawyers on a test designed specifically for humans. And that's just one of 700+ tasks where AI has now officially surpassed human performance.

The superintelligence conversation isn't some distant sci-fi scenario anymore. We're watching it happen in real time, from AlphaFold solving 50-year-old protein folding mysteries to Chinese surveillance systems identifying faces in crowds of 50,000 people with 95% accuracy. But here's what most coverage misses: understanding these capabilities helps us spot the real risks before they become problems.

James Caldwell breaks down what 700 human-level AI achievements actually mean for the next five years. You'll understand why computational costs still matter (GPT-4's training bill hit $100 million), how current limitations create unexpected bottlenecks, and which specific capabilities we should be watching most carefully.

In This Episode:

> Why beating humans at tests doesn't equal general intelligence

> The resource constraints that still govern AI development

> Which AI capabilities pose immediate vs theoretical risks

> How to think about regulation when technology moves this fast

Timestamps:

00:00 Introduction

02:15 The 700 tasks breakdown

04:30 Computational limits vs capabilities

07:20 Real vs imagined risks

09:45 What to watch next

The gap between AI hype and AI reality is shrinking fast. Follow Unboxed to stay ahead of what's actually happening, not what's being promised. New episodes drop multiple times daily because AI doesn't wait for anyone.

Learn more about your ad choices. Visit megaphone.fm/adchoices

Boston Dynamics just released footage of their Atlas robot doing something that shouldn't be possible: running full speed, jumping over obstacles, and landing like it's been doing parkour for years. Most people see this and think "cool robot tricks." The reality? This changes everything about how machines will move through our world.

These aren't remote-controlled toys. Atlas uses real-time computer vision and machine learning to navigate terrain that would challenge most humans. When you watch it sprint across uneven ground and leap over barriers without missing a step, you're watching AI solve one of robotics' hardest problems: dynamic movement in unpredictable environments.

In This Episode:

> How Boston Dynamics went from creating realistic animal simulations to building robots that outperform nature

> Why Atlas running at 5.6 mph matters more than Spot's $75,000 price tag

> What happens when these movement capabilities combine with large language models

> The real timeline for when these robots move from labs to loading docks

James Caldwell breaks down the technical breakthroughs that make this possible, from the machine learning algorithms processing sensor data in milliseconds to the engineering challenges of keeping a 200-pound robot stable while airborne. Plus, why the military applications everyone talks about might not be the biggest story here.

Timestamps:

00:00 Introduction

01:30 Boston Dynamics' 30-year journey from MIT spin-off

03:45 How Atlas actually "sees" and processes terrain

06:20 The AI breakthrough that changed everything

08:15 Real-world applications beyond the hype

10:30 What this means for the future of work

🤖 New AI developments drop daily on Unboxed. Hit follow to stay ahead of what's actually happening in artificial intelligence.

Learn more about your ad choices. Visit megaphone.fm/adchoices

Microsoft just revealed GPT-4 will process video, text, and images simultaneously. That $500 billion AI race everyone's talking about? It might already be over.

While Google and OpenAI have been building separate models for different types of content, Microsoft quietly developed something different. Their upcoming GPT-4 variant can watch a video, read the caption, analyze the thumbnail, and understand how all three pieces connect. Think about what that means for search, content creation, and basically every AI application we use today.

James Caldwell breaks down why this multimodal approach isn't just a technical upgrade. It's a fundamental shift that could make current AI models look like calculators. The early tests from Microsoft's Cosmos-1 model show something pretty remarkable: it doesn't just see objects in images, it understands context, relationships, and even implied meaning between visual elements.

In This Episode:

> How Microsoft's multimodal AI actually works (and why it's different)

> Why combining video, text, and image processing changes everything

> What this means for Google's AI strategy and the competition ahead

> Real examples of what these systems can do that current models can't

The timing here matters. While competitors focused on making their text models smarter, Microsoft built something that thinks more like humans do. We don't process information in separate silos - we combine what we see, read, and hear instantly. That's exactly what this new GPT-4 can do.

Timestamps:

00:00 Microsoft's multimodal breakthrough

02:30 How Cosmos-1 leads to GPT-4 video

05:15 Why this beats Google's approach

08:45 What developers can build with this

11:20 The competitive implications

This is the kind of AI development that changes entire industries overnight. Follow Unboxed for daily AI updates that actually matter - James drops multiple episodes each day because this stuff moves too fast to wait.

------

Keywords: artificial intelligence, technology news, ai developments, ai simplified

Learn more about your ad choices. Visit megaphone.fm/adchoices

Google just created a robot that can look at your messy kitchen and figure out how to bring you snacks without anyone teaching it that specific task. PaLM-E isn't just another chatbot with robot arms attached. It's the first AI that truly connects language understanding with visual perception to handle real-world situations.

While everyone's debating whether AI will replace jobs, Google quietly built something that changes the game entirely. PaLM-E has 562 billion parameters and can switch between different robot bodies while maintaining its intelligence. Think of it like a brain that works whether it's in a wheeled robot or a robot arm.

In This Episode:

> How PaLM-E combines vision and language processing in ways previous AI couldn't

> Real tests showing the robot handling tasks it was never trained for

> Why this approach could solve the biggest problem in robotics right now

> What happens when James tries to stump the system with complex requests

The most impressive part? PaLM-E can understand instructions like "bring me the rice chips from the drawer" and figure out what rice chips look like, where drawers typically are, and how to navigate around obstacles to complete the task. No pre-programming required.

This isn't about replacing human workers tomorrow. It's about creating AI that can actually function in the unpredictable real world instead of controlled lab environments.

Timestamps:

00:00 What makes PaLM-E different

02:30 Live testing with real tasks

05:15 The vision-language breakthrough explained

08:00 What this means for consumer robots

10:45 Why most robotics companies are missing this

James breaks down the technical details without the Silicon Valley hype. If you want to understand where AI robotics is actually heading, this episode cuts through the noise.

Follow Unboxed for daily AI breakdowns that actually matter. New episodes drop throughout the week.

Learn more about your ad choices. Visit megaphone.fm/adchoices

Meta just dropped a bomb on the AI world, and OpenAI executives are probably having some very uncomfortable meetings right now.

Here's what happened: Meta released LLaMA, a language model that's about to change everything we thought we knew about AI efficiency. The smallest version, LLaMA-13B with just 13 billion parameters, is outperforming GPT-3's 175 billion parameters on most benchmarks. That's like a Honda Civic beating a Ferrari in a race while using half the gas.

But it gets crazier. Within days of Meta releasing LLaMA to select researchers, someone leaked the entire thing on BitTorrent. Now anyone with a decent graphics card can run what was supposed to be cutting-edge, restricted AI technology from their bedroom.

In This Episode:

> Why LLaMA's efficiency breakthrough has AI companies scrambling to redesign their models

> The leaked BitTorrent files that democratized billion-parameter AI overnight

> What this means for the future of AI accessibility and who controls these tools

> How Meta trained LLaMA on 1.4 trillion tokens without breaking the bank

Felix breaks down the technical details that make LLaMA so efficient and why this leak might be the most significant moment in AI since ChatGPT launched. You'll understand exactly why OpenAI's business model just got a lot more complicated.

Timestamps:

00:00 The LLaMA leak that shocked Silicon Valley

02:30 Breaking down the efficiency numbers

05:45 Why this changes AI economics forever

08:15 The democratization vs safety debate

11:00 What happens next

If you want to understand AI moves before they hit mainstream tech news, follow Unboxed. Felix drops new episodes daily with the analysis that actually matters.

-------

Keywords: ai impact, ai ethics, midjourney, large language models, ai podcast, ai applications

Learn more about your ad choices. Visit megaphone.fm/adchoices

GPT-4 just scored in the 90th percentile on the bar exam. That's better than most actual lawyers.

OpenAI's latest model isn't just a text upgrade. It can analyze images, maintain longer conversations without losing its train of thought, and it's 40% more likely to give you accurate information. But here's what most people are missing: this isn't just about chatbots getting smarter. This is about AI crossing into professional-level reasoning.

Felix breaks down the numbers that actually matter. We're talking about a system that went from 10th percentile to 90th percentile on professional exams in one iteration. That's not incremental improvement. That's a fundamental shift in what AI can do.

In This Episode:

> Why GPT-4's image processing changes everything for data analysis

> The real implications of AI scoring better than 90% of lawyers

> How longer context windows affect business applications

> What this means for knowledge workers in 2024

You'll understand exactly why this update has tech leaders scrambling to rethink their AI strategies. Felix explains the technical improvements without the jargon, plus what it means for anyone whose job involves processing information.

Timestamps:

00:00 Introduction

02:30 GPT-4 vs GPT-3.5 performance breakdown

05:15 Image processing capabilities explained

07:45 Why context length matters more than you think

10:20 Job market implications and what's next

The AI race just accelerated. Don't get left behind wondering what happened. Follow Unboxed for daily episodes that keep you ahead of the curve without drowning you in technical complexity. Multiple new episodes drop daily.

---------

Keywords: ai applications, technology news, ai ethics, tech podcast, ai bias, tech explained, ai for beginners, artificial intelligence

Learn more about your ad choices. Visit megaphone.fm/adchoices

You're spending $50,000 a year on GPT-4 subscriptions across your team, but you're getting maybe 20% of its actual capabilities. Most users treat it like a slightly smarter Google search when it's actually a programmable reasoning engine.

Felix dug into the hidden features that OpenAI doesn't advertise and found some pretty shocking gaps in how people use GPT-4. That 100-message limit everyone complains about? There's a workaround. The 2000-word restriction that kills your longer prompts? Doesn't exist in the API. And the coding performance difference between 3.5 and 4 isn't just better, it's game-changing for complex builds.

The speed trade-off hits different when you know which tasks actually need GPT-4's horsepower versus what works fine on 3.5. Felix breaks down the cost-benefit math that most teams never calculate.

In This Episode:

> Why the ChatGPT interface limits GPT-4's true potential

> Prompt engineering techniques that actually work (not the Twitter guru nonsense)

> When to use 3.5 versus 4 based on real performance data

> How Chrome extension builders are leveraging model differences

> The hidden API features that change everything about workflow optimization

Timestamps:

00:00 The $50K revelation

02:15 Interface limitations nobody talks about

04:30 Prompt engineering that actually works

07:00 Speed versus quality trade-offs

09:20 API features you're missing

11:45 Wrap-up

Felix spent five years building ML models before Microsoft bought his startup, so he knows where the bodies are buried in AI development. His explanations cut through the hype to show you what actually matters.

Follow Unboxed for daily AI insights that won't waste your time. New episodes drop multiple times daily.

-----------

Keywords: machine learning, chatgpt, microsoft copilot, ai bias, ai news, algorithms, artificial intelligence, tech analysis

Learn more about your ad choices. Visit megaphone.fm/adchoices

Most people using Midjourney V5 are only scratching the surface. They type in basic prompts and wonder why their outputs look generic while others are creating stunning, professional-grade visuals that seem impossible to achieve with AI.

The problem isn't the technology. V5 actually processes images twice as fast as V4 and includes features that completely change what's possible. But these capabilities are buried in settings most users never touch and prompt techniques that aren't obvious from the interface.

Felix breaks down the specific tricks that separate amateur outputs from professional results. You'll learn why aspect ratio commands unlock cinematic possibilities, how the new tiling feature creates seamless textures without expensive software, and the image prompting workflow that lets you upload reference photos for precise variations.

In This Episode:

> Why V5's instant upscaling feature changes everything about iteration speed

> The aspect ratio hack that creates wide banner images and movie-style shots

> How to use tiling for seamless patterns that used to require Photoshop expertise

> Image prompting techniques for uploading references and getting exact variations

> The prompt structure that consistently produces professional-quality results

> Why most people's settings are actually working against them

Timestamps:

00:00 Introduction to V5's hidden potential

02:15 Aspect ratio commands that unlock new formats

04:30 Tiling feature for seamless textures

06:45 Image prompting workflow breakdown

08:20 Advanced prompt structures

10:15 Settings optimization

These aren't theoretical tips. Felix tested each technique extensively and shows you the exact prompts and settings that work. If you're serious about AI image generation, this episode will immediately upgrade your results.

Follow Unboxed for daily AI breakdowns that actually help you build better things. Felix drops new episodes every day, covering the tools and techniques that matter most for creators and builders.

-----

Keywords: ai simplified, microsoft copilot, automation

Learn more about your ad choices. Visit megaphone.fm/adchoices

Microsoft's new Copilot just turned every Office worker into a potential AI power user. And companies are scrambling to figure out what this means for their workforce.

This isn't just another AI assistant. Microsoft integrated GPT-4 technology directly into Word, Excel, PowerPoint, Outlook, and Teams. We're talking about AI that can analyze spreadsheets with thousands of rows in seconds, write entire reports from bullet points, and build presentations that actually look professional.

In This Episode:

> How Copilot reduces document creation time by 70% for routine work

> Why the $30 per month premium tier has executives doing math

> Real examples of what this AI can and can't handle in practice

> What this means for knowledge workers and office productivity

Robin breaks down the technical capabilities behind Microsoft's biggest Office update in decades. Early testing shows impressive results, but there are limitations most companies haven't considered yet. Some tasks that seem perfect for AI automation still need human oversight, while others that look complex actually work flawlessly.

The pricing strategy reveals Microsoft's confidence in this technology. Adding $30 monthly per user isn't cheap, but early adopters report time savings that justify the cost for certain roles. The question isn't whether this AI works, it's whether your company can afford not to use it.

Timestamps:

00:00 Microsoft's Copilot announcement breakdown

02:30 GPT-4 integration across Office apps

05:15 Real-world testing results and limitations

07:45 Pricing strategy and ROI calculations

10:20 What this means for different job roles

If you're trying to understand AI developments without the marketing fluff, follow Unboxed. Robin delivers multiple episodes daily breaking down what's actually happening in artificial intelligence.

----

Keywords: gpt-4, openai, ai podcast

Learn more about your ad choices. Visit megaphone.fm/adchoices