AI PULSE - How AI Deep Research Changes Everything

Affordable AI Revolution: Mistral Medium 3's High Performance at Lower Costs Introducing web search on the Anthropic API Google's Gemini API Offers 75% Cost Savings with Implicit Caching for AI Models Perplexity and Wiley Partner to Enhance Learning with AI-Driven Educational Content Integration Amazon's Vulcan Robot Revolutionizes Fulfillment Centers with Human-Like Dexterity Apple Explores AI Search Engines for Safari Amidst Google Antitrust Trial How AI Deep Research Changes Everything #AI, #ArtificialIntelligence, #Technology, #Innovation, #MachineLearning, #AIResearch, #TechNews

Episode Transcript

(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.
Mistral Medium 3 disrupts AI markets,
Anthropic Enhances real-time capabilities,
Google introduces cost-saving caching,
Perplexity partners with Wiley,
Amazon advances robotics,
and Apple explores AI search engines.
After this, we'll dive deep into AI deep research tools,

(00:33):
transforming information gathering.
Stay tuned.
The Mistral Medium 3 model is challenging the AI landscape
by offering near-high-end performance at a fraction of the cost.
Traditionally, AI deployment has been expensive,
requiring significant cloud resources or costly hardware.

(00:54):
Mistral Medium 3 changes this with a pricing model
of $0.40 per million input tokens
and $2 per million output tokens,
making AI more accessible.
The model is particularly adept at code generation and STEM tasks,
offering multimodal capabilities.
It's available across multiple platforms,

(01:15):
providing flexibility and avoiding vendor lock-in.
This approach democratizes AI,
enabling smaller companies to leverage advanced capabilities
and pushing larger providers to reconsider their pricing.
As part of Europe's technological push,
Mistral aims to remain independent,
highlighting the strategic importance of domestic AI development.

(01:38):
Mistral Medium 3 exemplifies a shift towards cost-efficient AI accessibility,
paving the way for broader adoption.
Let's now turn our attention to the integration potential,
Anthropoc has introduced Web Search on its API,
enhancing Cloud's ability to access real-time information from across the web.

(02:03):
This tool allows developers to build applications and agents
that deliver updated insights without needing separate search infrastructure.
Cloud can perform searches to provide accurate responses,
generating queries, retrieving and analyzing results with citations.
The tool supports industries like finance, legal and tech

(02:26):
by offering current data and specialized knowledge.
Web Search includes features for trust,
such as source citation and domain control,
ensuring information accuracy.
It integrates with Cloud Code for development workflows,
aiding in accessing the latest documentation and technical updates.
Companies like Quora's Poe and Adaptive.ai

(02:50):
provide the tools effectiveness in delivering fast, thorough results.
Web Search is available for Cloud version's 3.7 Sonnet and others,
priced at $10 per 1,000 searches.
Google is introducing implicit caching in its Gemini API,
promising up to 75% cost savings for third-party developers

(03:14):
using its Gemini 2.5 Pro and 2.5 Flash models.
This feature automatically caches frequently used data,
reducing computing requirements and costs.
Previously, Google's explicit caching required manual setup,
which some developers found cumbersome and costly.

(03:37):
Now, with implicit caching enabled by default,
savings occur automatically when a request shares a common prefix
with previous ones.
This system requires a minimum of 1,024 tokens for Flash
and 2,048 for Pro models.
Google advises placing repetitive context at the beginning of requests

(04:00):
to maximize cash-a-hits.
However, there's no third-party verification of these savings,
so developers remain cautious.
Early adopters will reveal its effectiveness in practice.
The Perplexity team has partnered with Wiley to integrate their answer engine
with Wiley's vast library of scientific, technical and medical materials.

(04:24):
This collaboration allows instructors and students to access Wiley resources
directly through Perplexity, enhancing learning efficiency and engagement.
Educational institutions benefit from streamlined curriculum delivery,
granting students quick access to Wiley's academic content.

(04:45):
Students can explore textbook materials and receive real-time explanations
and examples, focusing more on learning and less on searching.
This partnership aligns with Gen Z's learning preferences,
offering study guides, tailored explanations and real-world applications
without platform switching.
Educators gain tools for creating customized materials

(05:09):
and adapting lessons to current events.
Additionally, this initiative supports critical AI literacy,
teaching students to use AI responsibly, ensuring factual accuracy,
and fostering critical thinking skills vital for their future.
For now, let's focus on Vulcan's impact on safety.

(05:32):
When dropping a coin, humans effortlessly use sensors like hearing and touch to retrieve it,
a task difficult for most robots.
Traditional industrial robots often lack sensory feedback, leading to rigid responses.
At Amazon's Delivering the Future event, they introduced Vulcan,
a robot with a sense of touch, marking a significant advancement in robotics.

(05:58):
Vulcan is capable of delicate tasks, such as picking and stowing items in crowded bins,
thanks to its sensory technology.
This innovation enhances operations at fulfillment centers, improving safety and efficiency.
Vulcan works alongside employees, handling tasks like reaching high inventory rows,

(06:20):
reducing physical strain, the robot uses force feedback sensors
and physical AI to learn and improve its capabilities.
Amazon plans to deploy Vulcan widely, aiming to enhance operational efficiency
and workplace safety across their network.
Apple executive Eddie Q revealed that Apple is exploring AI search engines for Safari,

(06:45):
potentially challenging Google's dominance.
Q testified during a US Department of Justice antitrust trial against Google,
noting that Safari searches declined for the first time as users shifted to AI-based solutions.
Q stated that before AI, other search options weren't viable,
but now new entrants like Perplexity and OpenAI offer alternatives with chat-like interfaces.

(07:12):
Apple has already engaged in discussions with Perplexity.
Although AI search engines won't become the default due to their need for improvements,
they will be added to Safari's options.
Google's substantial payments to Apple for being the default might also influence this decision.
Apple plans to expand user choices beyond Google's Gemini,

(07:36):
while AI tools offer benefits and drawbacks,
balancing accurate responses with potential inaccuracies.
And now, pivot our discussion towards the main AI topic.
Today, we're going to explore the fascinating world of AI deep research features,

(07:59):
a powerful new capability that's transforming how we gather and process information.
These AI-powered research assistants are now available in several leading AI platforms,
including ChatGPT, Google Gemini, and most recently, Claude.
I'm your host, Fred, and joining me today is technology analyst, Yakov Lasker,

(08:22):
who has extensively tested and compared these tools.
Yakov, welcome to Innovation Pulse.
Thanks for that introduction, Fred.
I'm excited to dive into this topic with you today.
AI deep research is really revolutionizing how we approach information gathering.
Please go ahead with your first question.
Let's start with the basics.

(08:43):
Many of our listeners might be hearing about deep research for the first time.
What exactly is AI deep research, and how did it evolve?
At its core, deep research is an AI capability that acts as an autonomous research assistant.
Instead of you spending hours browsing dozens or hundreds of websites to research a topic,
these AI agents search the web on your behalf, analyze various sources,

(09:07):
and compile comprehensive reports with their findings.
The feature first appeared in Google's Gemini in December 2024,
making them pioneers in this category.
Open AI followed by introducing deep research to ChatGPT in February 2025,
and most recently in April 2025, Anthropik added the research feature to Claude.

(09:29):
That timeline is helpful.
I'm curious about how these systems actually work behind the scenes.
How do these deep research features operate differently from standard AI responses?
These tools fundamentally differ from standard AI responses in that they're more agintic,
meaning they operate with more autonomy.
When you submit a query, the AI transforms it into a research plan with multiple steps

(09:52):
then actively searches and browses the web to find relevant information.
The systems use multiple searches that build upon each other, determining what to investigate next.
They show their reasoning process as they work,
and ultimately synthesize all the gathered information into a comprehensive report with citations.
It's like having a research assistant who can read hundreds of web pages in minutes

(10:15):
and extract the key insights.
That's fascinating.
I notice you used the term agintic.
How does this concept of AI agency apply specifically to deep research features?
The agency in AI refers to the system's ability to take independent actions to achieve goals.
With deep research, the AI doesn't just answer with what it already knows.

(10:36):
It actively explores the web, makes decisions about which sources to trust,
and determines which information is relevant to your query.
For example, Claude's research feature conducts multiple simultaneous searches that build upon each other.
Gemini creates a multi-point research plan and shows its thinking as it reasons through information.
These systems are making judgment calls about what to explore next, much like a human researcher would,

(11:02):
but at a much faster pace and larger scale.
I see users now have options across multiple platforms.
What are the key differences between how ChatGPT, Gemini, and Claude implement their deep research features?
There are several important distinctions.
ChatGPT offers two versions, a full version that provides in-depth reports,

(11:23):
but takes up to 30 minutes to complete, and a lighter version that's faster but less thorough.
They limit the number of queries based on your subscription tier.
Gemini's deep research is available to both free and paid users, though with different query limits.
Notably, Google made this feature available for free users first, while it requires payment on other platforms.

(11:47):
Claude's research feature is the newest entrant, launching in April 2025.
Its main differentiator is speed.
While maintaining quality, it typically completes research in just a couple of minutes, compared to 7 to 10 minutes for competitors.
Claude also offers integration with Google Workspace for enterprise users.

(12:08):
That's a good rundown of the differences.
Speaking of real-world testing, you've put these tools through their paces.
Can you share what types of queries work best with deep research features?
Deep research excels with complex, multifaceted topics that benefit from synthesizing information from numerous sources.
Topics like market analysis, technology comparisons, scientific concepts, and historical developments work particularly well.

(12:36):
For instance, I've tested prompts like,
explore how time travel is portrayed in film and television,
or investigate the effectiveness of ketogenic diets for long-term weight loss.
Questions about consumer products also work great, like, evaluate the pros and cons of popular smart TVs.
However, these tools aren't ideal for simple factual queries that only need a quick answer,

(13:02):
or for highly specialized technical topics where accuracy is critical.
They're best for research questions that would otherwise require hours of manual reading.
So if I was looking to purchase a new laptop and wanted an overview of options within my budget, that would be an ideal use case.
Exactly, that's a perfect example.
Instead of visiting dozens of review sites and comparing specs yourself,

(13:25):
you could simply tell the AI your budget, preferred features, and use case.
A deep research feature would then browse numerous tech review sites, product pages,
and comparison articles to create a comprehensive report.
You'd get detailed information about the best options within your price range, pros and cons of each model,
and even user feedback patterns that might not be obvious from a quick search.

(13:49):
This saves hours of manual research while still giving you a balanced view to make an informed decision.
How reliable are the results from these deep research tools?
Do they suffer from the same issues as standard AI responses, like hallucinations or outdated information?
Reliability varies, and yes, these systems do sometimes struggle with similar issues as standard AI responses.

(14:11):
In my testing, I've found that hallucinations, making up information, can still occur,
particularly when dates and timelines are involved.
For instance, chat GPT's deep research has misidentified product release dates,
or considered upcoming products as already released.
Unmuted information can be a problem too.
In one test comparing AI chip companies, chat GPT relied heavily on sources from 2017 and 2021,

(14:37):
missing recent developments like Google's TPU V6e chip.
Gemini sometimes produces less detailed reports, but with more current information.
Always verify crucial information by checking the provided sources,
especially for time-sensitive topics or when making important decisions based on the research.
That's an important caveat.

(14:59):
Now, let's talk about accessibility.
What are the costs associated with using these deep research features across different platforms?
The pricing structures vary significantly.
Chat GPT's deep research requires a chat GPT plus subscription at $20 a month,
with limits on how many full research queries you can run, typically 10 per week.

(15:20):
Google Gemini offers deep research to both free and paid users.
Free users get 5-10 queries per month, while paid Gemini advanced users 20 month.
Get around 20 queries per day.
Claude's research feature initially launched for their premium tiers,
Claude Max, 100 month, team, and enterprise plans.

(15:42):
However, Anthropoc has stated they plan to make it available to Claude Pro subscribers 20 month in the future.
Currently, it's only available in the US, Japan, and Brazil.
So Gemini currently offers the best value with some free access, while the others require descriptions.
How do different AI platforms' deep research features handle specialized technical content?

(16:05):
Do any excel in particular domains?
Based on my testing and what others have reported, there are some differences in how they handle specialized content.
Chat GPT tends to produce more detailed, comprehensive reports for scientific and technical topics,
often including well-structured data tables and inline citations.

(16:26):
Gemini leverages Google's search expertise and often provides more current information,
particularly for rapidly evolving technical fields.
It excels at sourcing information from a larger number of websites,
sometimes analyzing over 100, 170 sites compared to Chat GPT's 20, 30 sources.
Claude's approach is often more concise and direct.

(16:50):
For instance, when asked about mahjong rules, while Chat GPT and Gemini included extensive historical context and background,
Claude focused on creating practical, step-by-step instructions that a beginner could immediately use.
None consistently outperforms the others across all technical domains,
so the best choice depends on your specific needs and topic.

(17:14):
That's interesting. They each seem to have their own strengths.
I'm curious about the output formats. How do these tools present their research findings?
The presentation styles differ noticeably.
Chat GPT typically delivers polished, well-structured reports with clear sections, occasional data tables,
and a conversational writing style that's engaging to read.

(17:37):
It often feels like a professionally written article.
Gemini tends toward a more academic approach with thorough explanations of concepts before diving into specifics.
It frequently includes helpful comparison charts and tables.
Some users describe its style as more verbose but comprehensive.
Claude's output is often described as bookish or thoughtful,

(18:00):
a balanced approach between practical information and deeper insights.
It tends to emphasize experimentation and application rather than just information delivery.
All three include citations, though they implement them differently.
Some inline, some as footnotes, and some as links to the original sources.
We've talked a lot about differences, but are there areas where all three deep research implementations fall short?

(18:25):
Absolutely. Despite their impressive capabilities, all three systems have similar limitations.
First, they all struggle with very recent information.
Anything from the past few days or weeks may not be captured accurately.
Second, they sometimes oversimplify complex topics, particularly when the nuances require specialized expertise.

(18:46):
They can present conventional wisdom without acknowledging significant debates in a field.
Third, all three occasionally produce hallucinations or factual errors that can be difficult to spot
without verifying against the original sources.
This is especially problematic for niche topics with limited, reliable online information.

(19:08):
Finally, they all have constraints on how many deep research queries you can run,
which limits their utility for extensive or ongoing research projects.
Looking toward the future, how do you see these deep research features evolving over the next year or two?
I expect several major developments.
First, we'll likely see significant speed improvements while maintaining quality.
Claude has already pushed in this direction by completing research in minutes rather than 10-plus minutes.

(19:35):
Second, multimodal capabilities will expand.
Right now, chat GPT can analyze images and PDFs as part of its research,
but this will become more sophisticated across all platforms,
potentially including video and audio content analysis.
Third, customization will improve, allowing users to specify preferred sources, research methodologies, and output formats.

(20:02):
We might see specialized research modes for academic, business, or technical domains.
Finally, I expect deeper integration with other tools and services.
Claude has already begun with Google Workspace Integration,
but we'll likely see connections to many more data sources and specialized databases in the future.

(20:24):
Those predictions make a lot of sense.
Let's shift to a more practical question.
For someone just starting to use these tools, do you have any tips to get the most out of deep research features?
Definitely. First, be specific in your queries.
Instead of asking, tell me about electric vehicles.
Try something like compare the range, charging infrastructure, and total cost of ownership for top 2025 electric SUVs under $50,000.

(20:53):
Second, request specific output formats when needed.
You can ask for tables, bullet points, or specific sections if that would make the information more useful for your purposes.
Third, always verify critical information by checking the provided sources, especially for time-sensitive or consequential decisions.

(21:14):
Finally, use follow-up questions.
These tools maintain context, so you can ask for clarification or deeper exploration of specific points from the initial research without starting over.
Those are excellent tips.
One aspect we haven't touched on is ethics.
What ethical considerations should users keep in mind when using these deep research tools?

(21:37):
Several important ethical considerations come to mind.
First, attribution and intellectual property.
These tools aggregate information from many sources, but users should still properly cite sources when using this research in their own work.
Second, bias awareness.
The tools can inadvertently amplify biases present in their training data or in the sources they consult.

(22:00):
Always approach results with critical thinking and seek diverse perspectives.
Third, overreliance.
While these tools are powerful, they shouldn't completely replace human judgment, especially for consequential decisions.
They're best used as assistance rather than final authorities.
Finally, there's the potential for misuse.

(22:21):
These tools could be used to generate convincing misinformation at scale.
Users have a responsibility to apply these capabilities ethically.
That's a thoughtful perspective on the ethical dimensions.
Now, I'm curious about the impact of deep research on traditional research methods.
How do you see these tools changing how we approach information gathering?
They're fundamentally changing the research process in several ways.

(22:45):
First, they're dramatically compressing the initial exploration phase.
What might take hours or days of preliminary reading can now be accomplished in minutes.
Second, they're democratizing access to synthesized information.
People without specialized research training can now get comprehensive overviews of complex topics quickly, lowering barriers to knowledge.

(23:06):
Third, they're shifting the researcher's role from information gathering to information validation and analysis.
The human's job becomes more about evaluating the AI's findings and diving deeper where needed.
However, I don't see them replacing traditional research entirely.
They're best viewed as powerful assistants that handle the initial heavy lifting,

(23:29):
allowing humans to focus their attention on critical analysis, novel connections, and creative applications of the information.
As we approach the end of our conversation, what have been the most surprising findings from your testing of these deep research features?
What surprised me most was how differently each system approached the same query.
Given identical prompts, they often emphasized completely different aspects of a topic or structured their responses in unique ways.

(23:55):
I was also impressed by the speed improvements we've seen.
When ChatGPT first launched deep research, it could take up to 30 minutes for a comprehensive report.
Now, Claude can deliver similar quality in just a couple of minutes.
That's remarkable progress in just a few months.
Perhaps most surprising was how useful these tools could be for niche topics.
I expected them to perform well for mainstream subjects, but they often uncovered valuable insights on specialized topics

(24:22):
that would have required extensive manual research to discover.
Those are fascinating observations.
For our final question, if you had to recommend just one of these deep research tools to our listeners, which would you choose and why?
That's a tough question because each has distinct advantages.
If I had to choose one today, I'd probably recommend Google Gemini's deep research for most users,

(24:46):
primarily because it offers the best balance of accessibility, quality, and cost.
The fact that Gemini offers free access to this feature, albeit with limits, means anyone can try it without financial commitment.
It consistently produces solid research results across a wide range of topics,
analyzes a larger number of sources than competitors, and typically completes research in a reasonable timeframe.

(25:10):
That said, if you're already paying for ChatGPT+, or plan to subscribe to Claude Pro when the feature becomes available there,
you'll likely be satisfied with either of those options as well.
The differences aren't dramatic enough to justify subscribing to a service solely for its deep research capability.
Thank you, Yaakov, for this comprehensive overview of AI deep research features.

(25:34):
You've given our listeners a lot to think about and experiment with.
To our audience, whether you're a student, professional, or just someone curious about the world,
these tools offer exciting new ways to explore information.
We hope you'll try them out and let us know your experiences.
That's all for today's episode of Innovation Pulse.
I'm Fred, signing off until next time.

(25:57):
Thanks for having me, Fred.
It's been a pleasure discussing this fascinating technology with you and your listeners.
Looking forward to seeing how these tools continue to evolve.
We've reached the end of today's podcast where we explored how innovations like Mistral Medium 3 and Vulcan Robot

(26:18):
are reshaping industries and discuss the transformative potential of deep research tools like ChatGPT, Google Gemini, and Claude.
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.

(26:39):
Stay tuned for more updates.