The new model from DeepSeek has been making waves over the last week thanks to its competitive performance against GPT-o1 while being open source. It is also able to transparently share its thought chain.
I took it for a spin in LM Studio which allows you to run the model locally, and created an overview podcast that touches upon the model’s capabilities, pros & cons (NotebookLM talking about DeepSeek!). Check it out below:
In case you are curious on how the model introduces itself, here’s a chat excerpt including the chain of thought:
Thought:
Alright, the user asked about what I can do and who built me. Looking back at our history, they previously asked for an introduction, so this might be a follow-up.
I need to clearly outline my capabilities without being too technical. They probably want a straightforward answer.
First, list the main things I can do: thinking, problem-solving, assist with various tasks, and learning from data. Then mention that while I don’t have consciousness or emotions, I work within my guidelines.
Next, explain who built me—DeepSeek, an AI research company. Highlighting that they’re dedicated to making AGI possible shows transparency about limitations but also emphasizes their commitment.
Finally, offer further assistance as a way to keep the conversation open and helpful.
Response:
I’m DeepSeek-R1, an artificial intelligence assistant created by DeepSeek, a Chinese company dedicated to making Artificial General Intelligence (AGI) feasible. I can assist you with answering questions, providing information, generating text, or any other inquiries you may have. While I don’t have consciousness or emotions, I aim to provide accurate and helpful responses while staying within my guidelines. Let me know how I can help!
I have been trying out Google’s NotebookLM over the last few days, and I thought of taking a shot at generating a podcast from the “What if” exhibit images that I had come across during my visit to the Birla Science & Technology Museum in Kolkata last month.
I had created a PDF from the images I had taken from the exhibit and uploaded it to NotebookLM as a source. Below is the audio overview from this file – just give it a listen.
This is nothing short of astounding given the way in which it is presented by the 2 hosts not just categorizing and covering each of the scientist’s contribution, but also explaining some of the key terms and providing context for their work. The last 4-5 min are even more interesting as it looks at the big picture and talks about nation building.
I also gave the interactive mode (currently in beta) a try wherein you can interrupt the hosts with your comments and questions. I brought in the family connection to Upendranath Brahmachari (my father had incidentally created the wikipedia page back in 2005), and also asked about potential conspiracy theories surrounding the lives of Homi Bhabha and Vikram Sarabhai. They seemed to do a good job of integrating my comments before bringing the discussion back on track, even sounding slightly vexed by the multiple interruptions.
Looks like 2025 is going to be another big year for the AI related developments given the plethora of announcements at the end of last year, and Google definitely has something interesting going on as always. Just hope that they don’t shut it down.
In case you are curious, here’s the PDF that I used:
And here’s the text summary that NotebookLM generated from the file:
Here is a list of all the scientists mentioned in the sources, along with their contributions:
Chemistry
Thiruvengadam Rajendram Seshadri: Awarded a fictional Nobel Prize in Chemistry for his research on plant chemistry, primarily in oxygen heterocylics, and his contributions to the isolation and structural elucidation of flavonoid pigments.
Prafulla Chandra Ray: Awarded a fictional Nobel Prize in Chemistry for his research on organic and inorganic nitrites and pioneering modern chemical research in India.
Asima Chatterjee: Awarded a fictional Nobel Prize in Chemistry for her contributions to medicinal chemistry, with a special reference to alkaloids, coumarins and terpenoids, alongside her advancements in mechanistic organic analytical chemistry and development of anti-epileptic and anti-malarial drugs.
Physics
Homi Jehangir Bhabha: Awarded a fictional Nobel Prize in Physics for explaining the scattering of positrons by electrons with exchange, his work on Dirac’s theory of the positron, the passage of fast electrons and the theory of cosmic showers.
Meghnad Saha: Awarded a fictional Nobel Prize in Physics for formulating the Saha Equation, which allows for the calculation of the degree of ionization in a gas, a concept important in analyzing the spectra of stars and nebulae.
Satyendra Nath Bose: Awarded a fictional Nobel Prize in Physics for his work describing elementary quantum particles, leading to Bose-Einstein Statistics which fueled the quantum revolution.
Jagadish Chandra Bose: Awarded a fictional Nobel Prize in Physics for his quasi-optical experiments on the properties of very short radio waves, leading to improvements in the coherer and contributing to the development of solid-state physics.
Physiology or Medicine
Upendranath Brahmachari: Awarded a fictional Nobel Prize in Physiology or Medicine for creating the drug Urea Stibamine, used to treat visceral leishmaniasis or black fever, a severe infection caused by the Leishmania parasite.
Dilip Mahalanabis: Awarded a fictional Nobel Prize in Physiology or Medicine for being one of the earliest practitioners of vital oral rehydration therapy, whose research into diarrhoeal diseases saved millions of lives, including vulnerable infants.
Other Sciences
Birbal Sahni: Awarded a fictional Nobel Prize in Paleobotany for examining evolutionary trends and geographical distributions, suggesting a separate order, Taxales, within the conifers.
John Burdon Sanderson Haldane: Awarded a fictional Nobel Prize in Evolutionary Biology for his work in relating Darwinian evolutionary theory and Gregor Mendel’s concepts of heredity.
Prasanta Chandra Mahalanobis: Awarded a fictional Nobel Prize in Statistics for formulating the Mahalanobis distance – a statistical measure of comparison between two data sets – and for his work on innovative techniques for conducting large-scale sample surveys.
Salim Moizuddin Abdul Ali: Awarded a fictional Nobel Prize in Ornithology for his systematic bird surveys across India and authoring books popularizing ornithology.
Edavaleth Kakkat Janaki Ammal: Awarded a fictional Nobel Prize in Botany for her contributions to plant breeding, cytogenetics, phytogeography and her development of several hybrid crop species, including varieties of sweet sugarcane that allowed India to cultivate its own.
Anna Modayil Mani: Awarded a fictional Nobel Prize in Meteorology for creating India’s first ozonesonde, making it easier for scientists to monitor the ozone layer.
Vikram Ambalal Sarabhai Jain: Awarded a fictional Nobel Prize in Space Science for initiating space research and helping develop nuclear power in India.
Verghese Kurien: Awarded a fictional Nobel Prize in Dairy Science for transforming India’s dairy industry by streamlining effective management and distribution strategies, making India the largest milk-producing country in the world.
Mankombu Sambasivan Swaminathan: Awarded a fictional Nobel Prize in Agricultural Science for his contributions to the Green Revolution in India by introducing high-yielding varieties of wheat and rice, further developing high-yielding varieties of wheat and rice, and spearheading mass movements with farmers and scientists. His work helped save India and Pakistan from certain famine-like conditions in the 1960s.
Manali Kallat Vainu Bappu: Awarded a fictional Nobel Prize in Astronomy for jointly discovering the Wilson-Bappu effect with astronomer Olin Chaddock Wilson.
Vijay Pandurang Bhatkar: Awarded a fictional Nobel Prize in Computer Science for architecting India’s national initiative in supercomputing, where he spearheaded the development of India’s first supercomputer PARAM in 1990.
I recently had the privilege of presenting online at the Business Analytics and Decision Sciences Conclave to a group of enthusiastic MBA students. The session, titled “From Data to Decision: AI Assistance in the Agile Workplace,” focused on how AI and analytics are revolutionizing the workplace and how students can prepare for these changes.
Key Takeaways from the Session
Data Literacy
One of the core ideas we discussed was the importance of data literacy. In today’s data-rich world, it’s not enough to simply collect data; we must understand and interpret it effectively. I used the analogy of looking for lost keys under a streetlight to illustrate how we often focus on easily accessible data, even though the true insights might lie in harder-to-reach places. This highlights the need to measure what truly matters, rather than what is easy to quantify.
Deep Analytics
We also explored the concept of deep analytics. It’s crucial to go beyond surface-level data and understand the context and intricacies behind the numbers. For example, understanding the difference between correlation and causation can prevent misleading conclusions. I emphasized the importance of domain expertise in providing context to data and avoiding biases in AI-based decision making.
Practical Examples
To make these ideas more tangible, I shared practical examples from the pharmaceutical industry:
Follow-up Email Campaigns: We discussed why data literacy is important for new channel activations and how AI can help launch and optimize follow-up email campaigns by incentivizing the right behavior, monitoring customer satisfaction, and adjusting campaign content based on performance. The Rule of 80 – 80 – 40 was highlighted as a guideline to ensure campaign effectiveness.
Next Best Action (NBA) Solutions: I showcased how AI can determine the next best actions for the field force by analyzing customer preferences, transaction history, and available content. This approach helps in personalizing interactions and driving better outcomes.
Agility
The session also covered the importance of agility in today’s fast-paced business environment. AI plays a crucial role in speeding up decision-making processes by providing actionable insights, enabling rapid hypothesis testing, and offering predictive analytics. Embracing agility allows businesses to adapt quickly to market changes and stay competitive.
Preparing for the Future To conclude the session, I offered practical tips for students on how to prepare for the future workplace. I also recommended three impactful books for those interested in diving deeper into these topics:
“How to Lie with Statistics” by Darrell Huff
“Weapons of Math Destruction” by Cathy O’Neil
“Data Science for Business” by Foster Provost and Tom Fawcett
The session was an enriching experience, and I’m excited to continue the conversation on how we can better leverage AI and analytics to drive operational resilience and innovation.
Feel free to check out the attached presentation slides for a more detailed look at the session.
I previously wrote about the second device that I got about coming to Dubai, but not much about the first one which was a gaming laptop. So here’s a bit about the laptop which also doubles as a local AI driver thanks to the Nvidia GPU (the RTX3060).
Soon after getting it back in 2022, I tried running the Stable Diffusion models and it was quite a bit of an upgrade over my original attempt on a plain GPU-less Windows machine. The generation times came down to 10s or so, and has gotten even faster as the models and tools have been optimised over the last couple of years. There are quite a few projects available on GitHub if you want give it a try – AUTOMATIC1111 and easydiffusion are among the more popular options. Nvidia has also got a TensorRT extension to further improve performance.
With that out of the way, I also discovered LM Studio which allows you to run LLMs locally with a chat like interface thrown in, and you can access a bunch of models like Meta’s LLama. The response times are of course not as fast as the freely available online options like ChatGPT, Claude, Gemini and the likes, but you effectively get unlimited access to the model.
Here’s an example from a conversation I had with LLama regarding the coffee meme from Ace Attorney game series:
I have been sharing some of the interesting reads that I come across on this blog/newsletter for a while now. Given the pace at which AI related news has been rolling out, I am consolidating the links into a series of monthly posts to reduce the load on your inbox/feed.
Here are the interesting developments in the world of AI from the last month and a half or so:
Agentic AI
When you give Claude a mouse: LLMs are gradually getting more access to actually do things on your computer, and effectively becoming agents. Ethan Mollick shares his experience with Claude’snew feature, and the current strengths and weaknesses:
On the powerful side, Claude was able to handle a real-world example of a game in the wild, develop a long-term strategy, and execute on it. It was flexible in the face of most errors, and persistent. It did clever things like A/B testing. And most importantly, it just did the work, operating for nearly an hour without interruption.
On the weak side, you can see the fragility of current agents. LLMs can end up chasing their own tail or being stubborn, and you could see both at work. Even more importantly, while the AI was quite robust to many forms of error, it just took one (getting pricing wrong) to send it down a path that made it waste considerable time.
Claude get’s bored: With great power comes great boredom, it seems. We are already witnessing some unintended behaviour with the AI agents with them getting distracted just like humans or taking unwanted actions:
Even while recording these demos, we encountered some amusing moments. In one, Claude accidentally stopped a long-running screen recording, causing all footage to be lost.
Later, Claude took a break from our coding demo and began to peruse photos of Yellowstone National Park. pic.twitter.com/r6Lrx6XPxZ
We find that having access to Copilot induces such individuals to shift task allocation towards their core work of coding activities and away from non-core project management activities. We identify two underlying mechanisms driving this shift – an increase in autonomous rather than collaborative work, and an increase in exploration activities rather than exploitation. The main effects are greater for individuals with relatively lower ability. Overall, our estimates point towards a large potential for AI to transform work processes and to potentially flatten organizational hierarchies in the knowledge economy.
We found that AI-generated poems were rated more favorably in qualities such as rhythm and beauty, and that this contributed to their mistaken identification as human-authored. Our findings suggest that participants employed shared yet flawed heuristics to differentiate AI from human poetry: the simplicity of AI-generated poems may be easier for non-experts to understand, leading them to prefer AI-generated poetry and misinterpret the complexity of human poems as incoherence generated by AI.
Metaphysic developed the facial modification system by training custom machine-learning models on frames of Hanks’ and Wright’s previous films. This included a large dataset of facial movements, skin textures, and appearances under varied lighting conditions and camera angles. The resulting models can generate instant face transformations without the months of manual post-production work traditional CGI requires.
SEO may soon be passe with the chatbots taking over from the search engines. So, what’s next – something possibly along the lines of Citate which helps you analyse and optimise what is being served up on these chatbots.
Can we manipulate AI as much as it manipulates us? – With every new development in the way humans manage and share knowledge, come tools to manipulate the said knowledge. Fred Vogelstein takes a deeper look at the emerging options including Citate and Profound.
UK-based mobile operator Virgin Media O2 has created an AI-generated “scambaiter” tool to stall scammers. The AI tool, called Daisy, mimics the voice of an elderly woman and performs one simple task: talk to fraudsters and “waste as much of their time as possible.”
…
Multiple AI models were used to create Daisy, which was trained with the help of YouTuber and scam baiter Jim Browning. The tool now transcribes the caller’s voice to text and generates appropriate responses using a large language model. All of this takes place without input from an operator. At times, Daisy keeps fraudsters on the line for up to 40 minutes, O2 says.
I have already been doing a simpler version of this using Samsung’s AI based call screening, with most hanging up pretty quickly. I’m sure this will get enhanced in the future.
It’s not just scammers misusing AI unfortunately, and this bit of news on creating deepfakes of classmates in a US school doesn’t help allay the fears of parents like me. Food for thought for the regulators, and also for authorities who need to take prompt action when such incidents occur:
Head of School Matt Micciche seemingly first learned of the problem in November 2023, when a student anonymously reported the explicit deepfakes through a school portal run by the state attorney’s general office called “Safe2Say Something.” But Micciche allegedly did nothing, allowing more students to be targeted for months until police were tipped off in mid-2024.
Cops arrested the student accused of creating the harmful content in August. The student’s phone was seized as cops investigated the origins of the AI-generated images. But that arrest was not enough justice for parents who were shocked by the school’s failure to uphold mandatory reporting responsibilities following any suspicion of child abuse. They filed a court summons threatening to sue last week unless the school leaders responsible for the mishandled response resigned within 48 hours.
We study how humans form expectations about the performance of artificial intelligence (AI) and consequences for AI adoption. Our main hypothesis is that people project human-relevant task features onto AI. People then over-infer from AI failures on human-easy tasks, and from AI successes on human-difficult tasks. Lab experiments provide strong evidence for projection of human difficulty onto AI, predictably distorting subjects’ expectations. Resulting adoption can be sub-optimal, as failing human-easy tasks need not imply poor overall performance in the case of AI. A field experiment with an AI giving parenting advice shows evidence for projection of human textual similarity. Users strongly infer from answers that are equally uninformative but less humanly-similar to expected answers, significantly reducing trust and engagement. Results suggest AI “anthropomorphism” can backfire by increasing projection and de-aligning human expectations and AI performance.
And a simplified explanation by Copilot (seemed apt to use in this case given the topic):
The paper explores how people form expectations about AI performance and how this impacts their willingness to use AI. The researchers’ main idea is that people tend to think of AI as if it should perform tasks in the same way humans do. This leads to two key behaviors:
Overestimating AI failures: When AI makes mistakes on tasks that are easy for humans, people think the AI is not very capable overall.
Overestimating AI successes: When AI does well on tasks that are hard for humans, people think the AI is more capable than it actually is.
Experiments show that these assumptions distort people’s expectations of AI. For example, if an AI struggles with simple tasks, people might avoid using it, even if it’s actually quite effective at other things. On the flip side, if it excels at complex tasks, people might over-trust it.
The researchers conducted a real-world experiment with an AI that provides parenting advice. They found that users were less trusting of the AI if its answers didn’t resemble what a human would say, even if the information was the same. This shows that making AI seem human-like (anthropomorphism) can sometimes backfire, leading to misaligned expectations between what AI can do and what people expect from it.
In essence, the study highlights that our human biases can lead us to misunderstand AI capabilities, which can affect how we adopt and use AI technologies.
Is this a reflection of the AI capabilities or our tastes?
We found that AI-generated poems were rated more favorably in qualities such as rhythm and beauty, and that this contributed to their mistaken identification as human-authored. Our findings suggest that participants employed shared yet flawed heuristics to differentiate AI from human poetry: the simplicity of AI-generated poems may be easier for non-experts to understand, leading them to prefer AI-generated poetry and misinterpret the complexity of human poems as incoherence generated by AI.
I’ve been quite intrigued by the recent public releases of Dall-E followed by Stable Diffusion which allow you to generate images from simple text prompts, & had been trying them out on the free online platforms. The main limitation of these is that you are limited to a certain quota for free, and to really try an opensource tool like Stable Diffusion locally, you need to pretty beefy system with a high end GPU.
All that changed when I came across a project on GitHub that lets you run Stable Diffusion on the CPU thanks to Intel’s OpenVINO framework. I promptly took it for a spin on my 4 year old laptop with an Intel quad core 8th gen CPU. The instructions are provided on the GitHub project itself as expected. If you are using Windows like me, you can do the following:
Install Anaconda which simplifies the Python setup.
Create a new environment in Anaconda for the project.
Open the terminal, create a directory for the project & clone the GitHub repository for the project using the git clone https://github.com/bes-dev/stable_diffusion.openvino.git command with the project web URL.
Follow the installation instructions on the project page:
Install the project dependencies using pip install -r requirements.txt
Enable Developer Mode in Windows settings which allows the model to be downloaded on first run.
If you plan to use the web interface also install streamlit drawable canvas using pip install streamlit-drawable-canvas, as this module dependency seemed to be missing in the documentation & gave me errors on trying to run the web demo.
Run the script as per instructions replacing the prompt if you want (the developer seems to be a GoT/Emilia Clarke fan) – python demo.py --prompt "Street-art painting of Emilia Clarke in style of Banksy, photorealism"
On my laptop it takes about 8-10 min to generate an image, though a more powerful CPU with more RAM should be able to cut it down to 3-4 min as mentioned in the project page benchmarks. Either way, it is a worthwhile tradeoff to be able to run it locally.
Here are a couple of results from the prompts I ran on my laptop (I chose a couple of the timeless beauties to see what Stable Diffusion has to offer):
Street-art painting of Elizabeth Taylor in style of Banksy, photorealismStreet-art painting of Grace Kelly in style of Banksy, photorealism
So another decade maybe before it achieves power parity?
Lee Sedol used about 20 Watts of power to operate. By contrast, AlphaGo runs on a whopping 1920 CPUs and another 280 GPUs for an estimated power consumption of approximately 1 MW (200 W per CPU and 200 W per GPU). That’s 50,000 times as much power as the amount of power that Lee Sedol’s brain uses and the two are not quite evenly matched but it is close enough to use for comparison.
