What is Happening with DeepSeek this week

If you are a consistent reader you found my newsletter from last week covering the emergence of DeepSeek. Sinc then there have been many developments regarding DeepSeek that I think it more than warrants another look into the news and happenings around DeepSeek.

First news to cover is whether DeepSeek is a copy of OpenAI. The method DeepSeek used to train their models is something called Distillation. In basic terms this mean they trained the model on outputs from Chatgpt. They used carefully written prompts and then measured the output and ‘guessed’ how the reasoning model came up with the response. This is Distillation.

It is hard to tell how they really built this model because even though DeepSeek is open source the training models were not released along with the model itself. It is also hard to tell what it really cost to train the model without knowing what the training data was.

Countries in Europe and Asia are currently banning the use of DeepSeek over privacy concerns. Italy and Taiwan are 2 such countries doing so. I am sure more will be following their lead - but whom I don’t know.

In the United States senator Josh Hawley has proposed jail time and a very significant fine for those who use and download DeepSeek. Currently he is being flooded by constituents who do like the idea so I think this is unlikely to happen. But given that other Chineses company - Byte Dance - is put into legal limbo over their status I can see the government being involved at some point. This is merely speculation at this point.

What I think this signals or highlights a continuing trend toward the cost of AI is going to zero. We already see this trend in the increasing power and skill of open source models. Just look at the recent benchmarks for Mistral Small. It is performing on par with Gpt-3o. AI will become a tool that entrpreneurs build a wrapper around.(See article at the end of this newsletter issue for more.)

If younwant to practice using DeepSeek you can do so with AnythingLLM and LMStudio. Both now give you access to DeepSeekR1 and can run locally on your machine. I have found that DeepSeek is a little better than other current models but not markedly so. Running the model locally means that your data is. ot sen to external servers.

Also it is possible to start building automations with N8N and DeepSeek.

AI for SuperNova Detection

In the vast expanse of our universe, stars are constantly being born and dying. Some of these stellar deaths result in spectacular explosions known as supernovae. These cosmic events have long fascinated astronomers, providing crucial insights into the life cycles of stars and the evolution of galaxies. Now, thanks to cutting-edge technology and artificial intelligence, we’re entering a new era of supernova discovery and classification.

The Zwicky Transient Facility (ZTF), a state-of-the-art observatory that scans the night sky with its wide-field camera, has recently hit a remarkable milestone. In December, the facility announced that it had officially classified more than 10,000 supernovae. This achievement marks a significant leap forward in our understanding of these celestial phenomena.

But what’s truly exciting is how they’ve managed to reach this milestone. In 2023, the ZTF received a major upgrade in the form of machine learning algorithms. An international team, led by researchers at Northwestern University, developed the Bright Transient Survey Bot (BTSbot). This innovative algorithm has revolutionized the way supernovae are detected, identified, and classified.

Trained on an impressive dataset of over 1.4 million historical images, including confirmed supernovae and temporarily flaring stars, BTSbot has proven to be an invaluable tool for astronomers. Since its implementation, it has discovered half of the brightest supernovae before human researchers could identify them. This not only saves countless hours of human effort but also allows for faster follow-up observations of these fleeting cosmic events.

The implications of this advancement are far-reaching. With BTSbot’s ability to rapidly process vast amounts of data, astronomers can now build larger, more comprehensive catalogs of supernovae. This expanded dataset is crucial for deriving meaningful insights about stellar evolution and the universe at large.

The integration of AI and machine learning promises to unlock even more cosmic secrets. The success of BTSbot at the ZTF is just the beginning. We can expect similar AI-driven tools to enhance other areas of astronomy, potentially leading to groundbreaking discoveries about our universe.

For astronomy enthusiasts and professionals alike, this is an exciting time. The marriage of traditional observational techniques with cutting-edge AI is ushering in a new golden age of cosmic exploration. As we gaze up at the night sky, we can take comfort in knowing that our understanding of the stars above is growing by leaps and bounds, thanks to the tireless work of both human researchers and their AI assistants.

To Help me make BrainScriblr better please think about answering just a short question. Follow the link here to take a brief survey.

Buy Me a Coffee

New Frontier of Tech Startups and AI Wrappers

In the rapidly evolving landscape of artificial intelligence, a new breed of companies has emerged: AI wrappers. These startups are building products and services on top of powerful AI models developed by tech giants like OpenAI, Anthropic, and Google. While this approach allows for quick innovation and market entry, it also presents unique challenges and risks that both entrepreneurs and investors need to understand.

What Are AI Wrappers?

AI wrapper companies are startups that leverage third-party AI technologies, particularly large language models (LLMs), as the core of their products or services. These companies don't develop the underlying AI models themselves but instead create user interfaces, specific applications, or tailored solutions that make these powerful AI tools accessible and useful for particular use cases or industries.

We can categorize AI wrappers into three levels based on their dependency on outsourced AI:

1. Level 1: Companies using AI to reduce costs or boost employee productivity.

2. Level 2: Businesses with some revenue-driving features based on outsourced AI.

3. Level 3: Startups entirely built on third-party AI, which forms the backbone of their product or service.

The Financial Tightrope

One of the primary challenges facing AI wrappers is the delicate balance of costs and pricing. The APIs provided by AI companies like OpenAI are not cheap, and these costs directly impact the wrapper company's bottom line. For Level 2 and Level 3 wrappers, these AI costs become part of their Cost of Goods Sold (COGS), significantly affecting their gross margins.

This cost structure forces many AI wrappers to adopt usage-based pricing models. While this aligns costs with revenue, it also makes it challenging to predict and manage cash flow. Moreover, if the underlying AI provider increases its prices, the wrapper company may be forced to pass these costs onto customers, potentially damaging relationships and competitiveness.

The Defensibility Dilemma

Perhaps the most significant challenge for AI wrappers is building a defensible business. Since these companies are essentially creating interfaces or applications on top of someone else's technology, their unique value proposition can be precarious. As the underlying AI models improve, features that once required significant engineering effort may become trivial to implement, potentially rendering some wrapper companies obsolete.

This vulnerability is particularly acute for Level 3 wrappers, whose entire business model relies on third-party AI. Investors are increasingly wary of these companies, recognizing that their long-term viability is closely tied to factors outside their control, such as the pricing and feature decisions of their AI providers.

Valuation Challenges

The unique position of AI wrappers in the tech ecosystem has led to new approaches in valuing these companies. A simplified formula might look like this:

Value = Utility x Degree of Abstraction x Difficulty of the Abstraction

In essence, a wrapper company's value lies in how well it abstracts the complexity of the underlying AI, making it useful for a specific purpose, and how difficult it is for competitors to replicate this abstraction. However, as AI capabilities advance, the "difficulty of abstraction" factor may decrease, potentially eroding the company's value over time.

Strategies for Success

Despite these challenges, some AI wrapper companies are finding success. Here are strategies that can help:

1. Focus on niche markets: By targeting specific industries or use cases, wrappers can build deeper expertise and more tailored solutions that are harder to replicate.

2. Develop proprietary data or models: While relying on third-party AI for core functionality, wrappers can develop their own specialized models or datasets to enhance their offerings.

3. Build strong network effects: Creating platforms that become more valuable as more users join can help create a moat around the business.

4. Diversify AI providers: Relying on multiple AI backends can reduce dependency on any single provider.

5. Invest in AI engineering: Continuously optimizing how the third-party AI is used can help manage costs and improve performance.

The Future Landscape

As AI technology continues to advance at a breakneck pace, the landscape for AI wrapper companies will undoubtedly evolve. We may see a consolidation in the market as some wrappers struggle to maintain relevance, while others may find innovative ways to add value beyond mere interface design.

Importantly, the relationship between AI providers and wrapper companies is likely to become more complex. AI giants may start viewing successful wrappers as potential acquisition targets or decide to enter their markets directly, further complicating the competitive landscape.

AI wrapper companies represent a fascinating new frontier in the tech startup world. They offer the promise of rapid innovation and the ability to bring powerful AI capabilities to diverse applications and industries. However, they also face unique challenges in building sustainable, defensible businesses.

For entrepreneurs venturing into this space, a clear-eyed understanding of these challenges is crucial. For investors, careful evaluation of a wrapper company's unique value proposition and long-term defensibility is essential. As the AI revolution continues to unfold, the story of AI wrappers will be one to watch closely, offering valuable lessons about innovation, business models, and the evolving relationship between technology providers and the ecosystems they enable.