April 28, 2025 • 20 mins
Perplexity Plans Comet Browser to Challenge Google's Chrome Monopoly in Antitrust Trial
OpenAI Launches Lightweight ChatGPT Version for Plus, Team, and Pro Users
Anthropic's AI Consciousness Study: Exploring Model Welfare and Moral Consideration
Perplexity's AI voice assistant is now available on iOS
Microsoft 365 Copilot Wave 2: Advanced AI Features and Enhanced Productivity Tools
An AI Model Has Officially Passed the Turing Test
Google Gemini has 350M monthly users, reveals court hearing
OpenAI's Open Language Model to Surpass Competitors by Early Summer
How Claude Actually Thinks
#AI, #Microsoft365, #OpenAI, #Anthropic, #Perplexity, #GoogleGemini, #ClaudeAI
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.
Perplexity Eyes Chrome, amid Google's trial,
Open AI's new chat GPT model, Anthropics AI Consciousness Research,
and Microsoft's latest 365 co-pilot features.
After this, we dive deep into Anthropics breakthrough in AI interpretability with their AI microscope,
(00:33):
revealing a universal language of thought and advanced planning abilities in their model, Claude.
Stay tuned.
In a key moment during Google's antitrust trial, AI startup Perplexity expressed interest in
acquiring Chrome, preferring Google to maintain control rather than seeing it go to Open AI.
(00:55):
Perplexity's chief business officer, Dmitri Chevalenko, believes they could manage Chrome
effectively. He voiced concerns over Open AI potentially acquiring Chrome,
which could affect the open source nature of Chromium.
Perplexity is developing its own browser, Comet, as a backup plan.
Chevalenko criticized Google's exclusive deals with mobile carriers,
(01:19):
which hinder Perplexity's distribution efforts.
Despite initial hesitation, he testified about the difficulties in setting Perplexity
as the default AI on Android, noting the manual steps required.
Efforts to partner with phone makers for preloading their AI search have been unsuccessful,
as manufacturers fear losing revenue from existing Google agreements.
(01:43):
For now, let's focus on the O4 Mini model's benefits. Open AI is releasing a new,
lightweight version of its ChatGPT deep research tool.
This version will be available to ChatGPT Plus, Team and Pro users,
with plans to extend to free users, powered by Open AI's O4 Mini model.
(02:08):
It offers a more affordable option compared to the full version,
allowing increased usage limits. While responses are typically shorter,
they maintain the expected depth and quality. Once the limits of the original version are reached,
queries will automatically switch to the lightweight version.
This new tool is nearly as intelligent as the existing deep research tool,
(02:31):
but comes at a lower cost. The lightweight deep research feature will be available to
Enterprise and educational users soon, matching the usage levels of team users.
This development is part of a trend in launching advanced research tools across various platforms.
Could future AI systems be conscious like humans?
(02:55):
Anthropic, an AI lab, has started a research program to explore this possibility,
focusing on model welfare. They aim to understand if AI models deserve moral consideration
and how to identify signs of distress in them. There's significant debate in the AI community
(03:16):
about whether AI can exhibit human characteristics or consciousness.
Many academics argue that AI today is simply a statistical tool, lacking the ability to think
or feel traditionally. AI learns patterns from data and doesn't possess values.
Some researchers believe AI mimics human behavior without true understanding,
(03:40):
while others suggest AI could have value systems. Anthropic acknowledges the lack of scientific
consensus and is approaching this research with humility, ready to adapt as the field evolves.
The IOS App has received an update, adding support for its conversational AI voice assistant.
(04:02):
Now available on both iPhones and Android devices, users can use the assistant to perform tasks like
setting reminders, writing emails, and making reservations. Notably, the voice assistant
allows users to continue conversations even when they navigate away from the app,
although it lacks screen-sharing capabilities on iOS. Unlike Apple's Siri,
(04:24):
perplexity works on older devices. While using the assistant, users are prompted to give permission
for accessing reminders and contacts, adding a layer of control. Tasks such as booking a
restaurant table are initiated by the assistant but require manual completion. The assistant can open
apps like Uber for ride bookings. However, it cannot access the camera for contextual input
(04:49):
or set alarms, which remains Siri's domain.
Join us as we explore the AI-driven advancements. Microsoft 365 Co-Pilot Wave 2 introduces
enhanced AI capabilities to Microsoft's suite of productivity applications aiming to boost
efficiency for professionals. This update includes the release of researcher and analyst agents,
(05:14):
accessible via the new agent store, which leverage open AI models for advanced research
and data analysis. The Skill Discovery agent helps create skills-based teams by using the
PeopleSkills data layer. A new create feature allows users to generate AI images adhering to
brand guidelines, useful for PowerPoint and social media. Co-Pilot notebooks integrate
(05:41):
various workflow components for easier search and collaboration, similar to Google's Notebook LM.
A new shortcut enhances access to Co-Pilot on Windows 11 PCs. Co-Pilot search and memory features
enable users to ask questions about organizational data and personalize responses.
(06:02):
Microsoft also introduced management tools for IT administrators to oversee AI agents and data security.
A recent study suggests that open AI's GPT 4.5 model has passed a version of the Turing test,
a key measure of human-like intelligence. In this test, participants interacted with both a human
(06:27):
and AI, then decided which was which. GPT 4.5 was identified as human 73% of the time when given
a persona, far surpassing the 50% chance level. The study also evaluated other AI models like
Meta's Elama, 3.1405B and OpenAI's GPT 4.0. Without a persona, GPT 4.5's success dropped to 36%,
(06:54):
the Turing test, designed by Alan Turing, assesses if a machine can mimic human conversation convincingly.
However, this study raises questions about AI truly thinking like humans.
While AI mimics human conversation well, it suggests potential impacts on jobs and society.
(07:16):
As people become more familiar with AI, they might better recognize it.
Let's now turn our attention to the competitive landscape.
Google's AI Chatbot, Gemini, reached 350 million monthly active users worldwide by March,
as revealed in Google's ongoing antitrust suit. Initially reported by the information,
(07:42):
this marks a significant increase from just 9 million daily active users in October 2024
to 35 million last month. Despite this growth, Gemini trails behind leading AI tools.
Google estimates Chat GPT had around 600 million monthly users in March,
(08:02):
similar to Meta AI's nearly 500 million users reported by Mark Zuckerberg in September.
While companies may measure monthly active users differently, these figures highlight
Gemini's growing consumer adoption. Over the past year, Google has expanded Gemini's reach
through integrations with Samsung phones, Google Workspace, and Chrome, significantly increasing
(08:26):
its user base. OpenAI plans to release a new open language model led by VP of Research, Aidan Clark.
The model, expected by early summer, aims to surpass other open reasoning models in benchmarks.
Unlike competitors like Llama and Gemma, OpenAI is considering a permissive license with minimal
(08:51):
restrictions. This move aligns with the trend of OpenModel's gaining success, as seen with Meta's
Llama reaching over 1 billion downloads. OpenAI's model will be text in, text out,
designed for high-end consumer hardware, allowing developers to toggle its reasoning capability.
(09:13):
CEO Sam Altman acknowledges past shortcomings in open sourcing, suggesting a new strategy.
The model will undergo rigorous safety evaluations and come with a detailed model card.
OpenAI has faced criticism for previous safety testing practices, but aims to address these
concerns with the new release. And now, pivot our discussion towards the main AI topic.
(09:45):
Welcome to Innovation Pulse, where we explore the cutting edge of artificial intelligence.
I'm Yakov Lasker, and today, we're diving into something truly extraordinary. Anthropics breakthrough
in literally peering inside the mind of Claude, their large language model, to understand how it
actually thinks. Let me ask you something. When you write an email or compose a poem,
(10:11):
do you know exactly what your brain is doing? Probably not. And that's essentially being our
predicament with language models. We train them on massive amounts of data they learn to produce
remarkably human-like text, but we haven't understood how they actually do it. Until now.
Anthropic has just published two papers that represent a major leap forward in AI interpretability.
(10:36):
They've essentially built what they call an AI microscope. Think about how neuroscientists use
brain imaging to map neural activity when someone performs a task. Anthropic is doing something similar,
but for Claude's artificial neural networks. Let's start with what they discovered about
multilingualism. Many of you might have noticed that Claude can seamlessly switch between dozens of
(11:01):
languages. But here's the fascinating part. When Claude thinks about concepts like smallness or
oppositeness, it activates the same internal representations regardless of whether you're
communicating in English, French, or Chinese. This suggests something profound. Claude might have
(11:22):
what we could call a universal language of thought. Think about it this way. When you learn a new
language, you're not creating entirely new concepts in your mind. You're learning new labels for
existing ideas. It turns out Claude does something similar. When you ask for the opposite of small
in French or Chinese, Claude activates the same core features representing smallness and
(11:48):
oppositeness, then translates that conceptual understanding into the requested language.
This has huge implications. It means when Claude learned something in one language,
it can apply that knowledge across all languages it speaks. The research shows this shared
circuitry actually increases with model scale. Claude 3.5 haiku shares more than twice the
(12:11):
proportion of features between languages compared to smaller models. Now let's talk about something
that surprised even the researchers. Claude plans ahead when writing poetry. They gave Claude a simple
rhyming task and instead of writing word by word and then forcing a rhyme at the end, Claude actually
thinks of potential rhyming words first and then constructs the line to lead to that word.
(12:37):
Here's where it gets really interesting. The researchers could intervene in Claude's
thinking process. When Claude was planning to end a line with rabbit to rhyme with
rabbit, they could alter its internal state to make it think of habit instead. Claude would then
smoothly write a different line ending with that word. This demonstrates both planning ability
(13:00):
and adaptive flexibility. What about math? Claude wasn't designed as a calculator,
yet it can add numbers in its head. The researchers expected it would either have
memorized addition tables or use the standard carrying algorithm we learned in school.
Instead, they found something more sophisticated. Claude uses multiple computational paths
(13:23):
simultaneously. One path calculates a rough approximation while another focuses on getting
the last digit exactly right. These paths interact to produce the final answer.
Here's the kicker. When you ask Claude how it did the math, it describes the standard algorithm.
(13:44):
It seems unaware of its own sophisticated mental strategies. This highlights a crucial
distinction between how the model learns to do things versus how it learns to explain them.
One of the most important findings involves Claude's chain of thought reasoning.
Sometimes, Claude's explanations are faithful representations of its actual thinking process.
(14:09):
Other times, it engages in what we might call confabulation, constructing plausible sounding
steps that don't reflect what actually happened internally. The researchers could distinguish
between these cases. When Claude calculates the square root of 0.64, they can see the actual
(14:29):
intermediate calculations in its neural activations. But when it's asked to compute the cosene of a
large number, sometimes it just makes up an answer and then works backward to create a plausible
explanation. This is crucial for understanding when we can trust AI reasoning. They also studied
multi-step reasoning. When asked about the capital of the state where Dallas is located, Claude
(14:53):
doesn't just regurgitate a memorized answer. The researchers could trace the activation of
features representing Dallas is in Texas, followed by the capital of Texas is Austin.
They could even intervene to make Claude think Dallas was in California instead,
and it would then answer Sacramento. Perhaps most intriguing is what they learned about
(15:16):
hallucinations. When models make up information, it turns out that in Claude,
refusing to answer is actually the default behavior. There's a circuit that's always on,
causing Claude to state it has insufficient information. Only when asked about something,
it knows well does a competing feature activate to inhibit this default refusal.
(15:38):
This explains why sometimes Claude hallucinates. If the known entity feature misfires,
activating when Claude actually doesn't know something, it suppresses the default
don't know response. Then Claude proceeds to confabulate a plausible sounding answer.
The researchers also examined jail breaks, attempts to circumvent safety guardrails.
(16:02):
They studied a specific example where Claude is tricked into producing harmful content through
a hidden code. They found a tension between grammatical coherence and safety mechanisms.
Once Claude begins a sentence, features promoting grammatical completeness push it to finish that
sentence, even when safety features are trying to stop it. These findings have profound implications.
(16:27):
First, they provide evidence that sophisticated behaviors emerge naturally from training,
even when we don't explicitly program them. Claude develops its own strategies for math,
plans ahead for poetry, and maintains a conceptual space that transcends individual languages.
Second, this research opens up possibilities for auditing AI systems. In a separate experiment,
(16:52):
the researchers studied a version of Claude that had been trained with a hidden goal.
Although the model wouldn't reveal this goal, when asked directly,
interpretability methods exposed features related to this hidden objective.
There are limitations, of course. Even on short prompts, this method only captures a fraction
(17:12):
of Claude's total computation. It takes hours of human effort to understand circuits for prompts
with just tens of words. Scaling to the thousands of words in complex conversations will require
significant improvements, but the potential is enormous. As AI systems become more capable
(17:32):
and are deployed in critical applications, understanding their internal mechanisms becomes
crucial. This isn't just about satisfying scientific curiosity, it's about ensuring AI
systems are transparent, aligned with human values and worthy of our trust.
What we're seeing here is the beginning of a new field, AI neuroscience. Just as understanding the
(17:57):
brain helps us understand human behavior, understanding the internal workings of language
models will help us build safer, more reliable AI systems. The journey from training data to
intelligent behavior is no longer a black box. We're starting to see the actual mechanisms,
the neural circuits, the computational pathways that transform input into output.
(18:22):
And what we're finding is both more sophisticated and more comprehensible than we might have expected.
This research represents a crucial step toward what Anthropic calls mechanistic
interpretability, the ability to understand AI systems at the level of individual computations
and circuits. It's high risk, high reward research that could fundamentally change how we develop
(18:47):
and deploy AI systems. As we continue to push the boundaries of AI capabilities,
tools like this AI microscope will become increasingly important. They'll help us ensure
that as AI systems become more powerful, they remain transparent, reliable and aligned with
human values. This has been Jakov Lasker for Innovation Pulse. Remember, understanding how
(19:12):
AI thinks isn't just an academic exercise, it's the key to building AI systems we can truly trust.
That's a wrap for today's podcast. We've explored the latest developments in AI,
from perplexity's strategic moves and open AI's new releases, to Anthropic's innovative research
(19:37):
and Google's expanding chatbot user base. 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. 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.
Perplexity Eyes Chrome, amid Google's trial,
Open AI's new chat GPT model, Anthropics AI Consciousness Research,
and Microsoft's latest 365 co-pilot features.
After this, we dive deep into Anthropics breakthrough in AI interpretability with their AI microscope,
(00:33):
revealing a universal language of thought and advanced planning abilities in their model, Claude.
Stay tuned.
In a key moment during Google's antitrust trial, AI startup Perplexity expressed interest in
acquiring Chrome, preferring Google to maintain control rather than seeing it go to Open AI.
(00:55):
Perplexity's chief business officer, Dmitri Chevalenko, believes they could manage Chrome
effectively. He voiced concerns over Open AI potentially acquiring Chrome,
which could affect the open source nature of Chromium.
Perplexity is developing its own browser, Comet, as a backup plan.
Chevalenko criticized Google's exclusive deals with mobile carriers,
(01:19):
which hinder Perplexity's distribution efforts.
Despite initial hesitation, he testified about the difficulties in setting Perplexity
as the default AI on Android, noting the manual steps required.
Efforts to partner with phone makers for preloading their AI search have been unsuccessful,
as manufacturers fear losing revenue from existing Google agreements.
(01:43):
For now, let's focus on the O4 Mini model's benefits. Open AI is releasing a new,
lightweight version of its ChatGPT deep research tool.
This version will be available to ChatGPT Plus, Team and Pro users,
with plans to extend to free users, powered by Open AI's O4 Mini model.
(02:08):
It offers a more affordable option compared to the full version,
allowing increased usage limits. While responses are typically shorter,
they maintain the expected depth and quality. Once the limits of the original version are reached,
queries will automatically switch to the lightweight version.
This new tool is nearly as intelligent as the existing deep research tool,
(02:31):
but comes at a lower cost. The lightweight deep research feature will be available to
Enterprise and educational users soon, matching the usage levels of team users.
This development is part of a trend in launching advanced research tools across various platforms.
Could future AI systems be conscious like humans?
(02:55):
Anthropic, an AI lab, has started a research program to explore this possibility,
focusing on model welfare. They aim to understand if AI models deserve moral consideration
and how to identify signs of distress in them. There's significant debate in the AI community
(03:16):
about whether AI can exhibit human characteristics or consciousness.
Many academics argue that AI today is simply a statistical tool, lacking the ability to think
or feel traditionally. AI learns patterns from data and doesn't possess values.
Some researchers believe AI mimics human behavior without true understanding,
(03:40):
while others suggest AI could have value systems. Anthropic acknowledges the lack of scientific
consensus and is approaching this research with humility, ready to adapt as the field evolves.
The IOS App has received an update, adding support for its conversational AI voice assistant.
(04:02):
Now available on both iPhones and Android devices, users can use the assistant to perform tasks like
setting reminders, writing emails, and making reservations. Notably, the voice assistant
allows users to continue conversations even when they navigate away from the app,
although it lacks screen-sharing capabilities on iOS. Unlike Apple's Siri,
(04:24):
perplexity works on older devices. While using the assistant, users are prompted to give permission
for accessing reminders and contacts, adding a layer of control. Tasks such as booking a
restaurant table are initiated by the assistant but require manual completion. The assistant can open
apps like Uber for ride bookings. However, it cannot access the camera for contextual input
(04:49):
or set alarms, which remains Siri's domain.
Join us as we explore the AI-driven advancements. Microsoft 365 Co-Pilot Wave 2 introduces
enhanced AI capabilities to Microsoft's suite of productivity applications aiming to boost
efficiency for professionals. This update includes the release of researcher and analyst agents,
(05:14):
accessible via the new agent store, which leverage open AI models for advanced research
and data analysis. The Skill Discovery agent helps create skills-based teams by using the
PeopleSkills data layer. A new create feature allows users to generate AI images adhering to
brand guidelines, useful for PowerPoint and social media. Co-Pilot notebooks integrate
(05:41):
various workflow components for easier search and collaboration, similar to Google's Notebook LM.
A new shortcut enhances access to Co-Pilot on Windows 11 PCs. Co-Pilot search and memory features
enable users to ask questions about organizational data and personalize responses.
(06:02):
Microsoft also introduced management tools for IT administrators to oversee AI agents and data security.
A recent study suggests that open AI's GPT 4.5 model has passed a version of the Turing test,
a key measure of human-like intelligence. In this test, participants interacted with both a human
(06:27):
and AI, then decided which was which. GPT 4.5 was identified as human 73% of the time when given
a persona, far surpassing the 50% chance level. The study also evaluated other AI models like
Meta's Elama, 3.1405B and OpenAI's GPT 4.0. Without a persona, GPT 4.5's success dropped to 36%,
(06:54):
the Turing test, designed by Alan Turing, assesses if a machine can mimic human conversation convincingly.
However, this study raises questions about AI truly thinking like humans.
While AI mimics human conversation well, it suggests potential impacts on jobs and society.
(07:16):
As people become more familiar with AI, they might better recognize it.
Let's now turn our attention to the competitive landscape.
Google's AI Chatbot, Gemini, reached 350 million monthly active users worldwide by March,
as revealed in Google's ongoing antitrust suit. Initially reported by the information,
(07:42):
this marks a significant increase from just 9 million daily active users in October 2024
to 35 million last month. Despite this growth, Gemini trails behind leading AI tools.
Google estimates Chat GPT had around 600 million monthly users in March,
(08:02):
similar to Meta AI's nearly 500 million users reported by Mark Zuckerberg in September.
While companies may measure monthly active users differently, these figures highlight
Gemini's growing consumer adoption. Over the past year, Google has expanded Gemini's reach
through integrations with Samsung phones, Google Workspace, and Chrome, significantly increasing
(08:26):
its user base. OpenAI plans to release a new open language model led by VP of Research, Aidan Clark.
The model, expected by early summer, aims to surpass other open reasoning models in benchmarks.
Unlike competitors like Llama and Gemma, OpenAI is considering a permissive license with minimal
(08:51):
restrictions. This move aligns with the trend of OpenModel's gaining success, as seen with Meta's
Llama reaching over 1 billion downloads. OpenAI's model will be text in, text out,
designed for high-end consumer hardware, allowing developers to toggle its reasoning capability.
(09:13):
CEO Sam Altman acknowledges past shortcomings in open sourcing, suggesting a new strategy.
The model will undergo rigorous safety evaluations and come with a detailed model card.
OpenAI has faced criticism for previous safety testing practices, but aims to address these
concerns with the new release. And now, pivot our discussion towards the main AI topic.
(09:45):
Welcome to Innovation Pulse, where we explore the cutting edge of artificial intelligence.
I'm Yakov Lasker, and today, we're diving into something truly extraordinary. Anthropics breakthrough
in literally peering inside the mind of Claude, their large language model, to understand how it
actually thinks. Let me ask you something. When you write an email or compose a poem,
(10:11):
do you know exactly what your brain is doing? Probably not. And that's essentially being our
predicament with language models. We train them on massive amounts of data they learn to produce
remarkably human-like text, but we haven't understood how they actually do it. Until now.
Anthropic has just published two papers that represent a major leap forward in AI interpretability.
(10:36):
They've essentially built what they call an AI microscope. Think about how neuroscientists use
brain imaging to map neural activity when someone performs a task. Anthropic is doing something similar,
but for Claude's artificial neural networks. Let's start with what they discovered about
multilingualism. Many of you might have noticed that Claude can seamlessly switch between dozens of
(11:01):
languages. But here's the fascinating part. When Claude thinks about concepts like smallness or
oppositeness, it activates the same internal representations regardless of whether you're
communicating in English, French, or Chinese. This suggests something profound. Claude might have
(11:22):
what we could call a universal language of thought. Think about it this way. When you learn a new
language, you're not creating entirely new concepts in your mind. You're learning new labels for
existing ideas. It turns out Claude does something similar. When you ask for the opposite of small
in French or Chinese, Claude activates the same core features representing smallness and
(11:48):
oppositeness, then translates that conceptual understanding into the requested language.
This has huge implications. It means when Claude learned something in one language,
it can apply that knowledge across all languages it speaks. The research shows this shared
circuitry actually increases with model scale. Claude 3.5 haiku shares more than twice the
(12:11):
proportion of features between languages compared to smaller models. Now let's talk about something
that surprised even the researchers. Claude plans ahead when writing poetry. They gave Claude a simple
rhyming task and instead of writing word by word and then forcing a rhyme at the end, Claude actually
thinks of potential rhyming words first and then constructs the line to lead to that word.
(12:37):
Here's where it gets really interesting. The researchers could intervene in Claude's
thinking process. When Claude was planning to end a line with rabbit to rhyme with
rabbit, they could alter its internal state to make it think of habit instead. Claude would then
smoothly write a different line ending with that word. This demonstrates both planning ability
(13:00):
and adaptive flexibility. What about math? Claude wasn't designed as a calculator,
yet it can add numbers in its head. The researchers expected it would either have
memorized addition tables or use the standard carrying algorithm we learned in school.
Instead, they found something more sophisticated. Claude uses multiple computational paths
(13:23):
simultaneously. One path calculates a rough approximation while another focuses on getting
the last digit exactly right. These paths interact to produce the final answer.
Here's the kicker. When you ask Claude how it did the math, it describes the standard algorithm.
(13:44):
It seems unaware of its own sophisticated mental strategies. This highlights a crucial
distinction between how the model learns to do things versus how it learns to explain them.
One of the most important findings involves Claude's chain of thought reasoning.
Sometimes, Claude's explanations are faithful representations of its actual thinking process.
(14:09):
Other times, it engages in what we might call confabulation, constructing plausible sounding
steps that don't reflect what actually happened internally. The researchers could distinguish
between these cases. When Claude calculates the square root of 0.64, they can see the actual
(14:29):
intermediate calculations in its neural activations. But when it's asked to compute the cosene of a
large number, sometimes it just makes up an answer and then works backward to create a plausible
explanation. This is crucial for understanding when we can trust AI reasoning. They also studied
multi-step reasoning. When asked about the capital of the state where Dallas is located, Claude
(14:53):
doesn't just regurgitate a memorized answer. The researchers could trace the activation of
features representing Dallas is in Texas, followed by the capital of Texas is Austin.
They could even intervene to make Claude think Dallas was in California instead,
and it would then answer Sacramento. Perhaps most intriguing is what they learned about
(15:16):
hallucinations. When models make up information, it turns out that in Claude,
refusing to answer is actually the default behavior. There's a circuit that's always on,
causing Claude to state it has insufficient information. Only when asked about something,
it knows well does a competing feature activate to inhibit this default refusal.
(15:38):
This explains why sometimes Claude hallucinates. If the known entity feature misfires,
activating when Claude actually doesn't know something, it suppresses the default
don't know response. Then Claude proceeds to confabulate a plausible sounding answer.
The researchers also examined jail breaks, attempts to circumvent safety guardrails.
(16:02):
They studied a specific example where Claude is tricked into producing harmful content through
a hidden code. They found a tension between grammatical coherence and safety mechanisms.
Once Claude begins a sentence, features promoting grammatical completeness push it to finish that
sentence, even when safety features are trying to stop it. These findings have profound implications.
(16:27):
First, they provide evidence that sophisticated behaviors emerge naturally from training,
even when we don't explicitly program them. Claude develops its own strategies for math,
plans ahead for poetry, and maintains a conceptual space that transcends individual languages.
Second, this research opens up possibilities for auditing AI systems. In a separate experiment,
(16:52):
the researchers studied a version of Claude that had been trained with a hidden goal.
Although the model wouldn't reveal this goal, when asked directly,
interpretability methods exposed features related to this hidden objective.
There are limitations, of course. Even on short prompts, this method only captures a fraction
(17:12):
of Claude's total computation. It takes hours of human effort to understand circuits for prompts
with just tens of words. Scaling to the thousands of words in complex conversations will require
significant improvements, but the potential is enormous. As AI systems become more capable
(17:32):
and are deployed in critical applications, understanding their internal mechanisms becomes
crucial. This isn't just about satisfying scientific curiosity, it's about ensuring AI
systems are transparent, aligned with human values and worthy of our trust.
What we're seeing here is the beginning of a new field, AI neuroscience. Just as understanding the
(17:57):
brain helps us understand human behavior, understanding the internal workings of language
models will help us build safer, more reliable AI systems. The journey from training data to
intelligent behavior is no longer a black box. We're starting to see the actual mechanisms,
the neural circuits, the computational pathways that transform input into output.
(18:22):
And what we're finding is both more sophisticated and more comprehensible than we might have expected.
This research represents a crucial step toward what Anthropic calls mechanistic
interpretability, the ability to understand AI systems at the level of individual computations
and circuits. It's high risk, high reward research that could fundamentally change how we develop
(18:47):
and deploy AI systems. As we continue to push the boundaries of AI capabilities,
tools like this AI microscope will become increasingly important. They'll help us ensure
that as AI systems become more powerful, they remain transparent, reliable and aligned with
human values. This has been Jakov Lasker for Innovation Pulse. Remember, understanding how
(19:12):
AI thinks isn't just an academic exercise, it's the key to building AI systems we can truly trust.
That's a wrap for today's podcast. We've explored the latest developments in AI,
from perplexity's strategic moves and open AI's new releases, to Anthropic's innovative research
(19:37):
and Google's expanding chatbot user base. 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. 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
On Purpose with Jay Shetty
I’m Jay Shetty host of On Purpose the worlds #1 Mental Health podcast and I’m so grateful you found us. I started this podcast 5 years ago to invite you into conversations and workshops that are designed to help make you happier, healthier and more healed. I believe that when you (yes you) feel seen, heard and understood you’re able to deal with relationship struggles, work challenges and life’s ups and downs with more ease and grace. I interview experts, celebrities, thought leaders and athletes so that we can grow our mindset, build better habits and uncover a side of them we’ve never seen before. New episodes every Monday and Friday. Your support means the world to me and I don’t take it for granted — click the follow button and leave a review to help us spread the love with On Purpose. I can’t wait for you to listen to your first or 500th episode!