Hi everyone

I have just posted in "Library" the original paper of DeepSeek V3 along with a summary and some definition of key concepts to fully understand the key takeaways of the paper. You can  also find the paper from the original github source.

For those new to LLMs or just starting to explore them, this video provides a straightforward and accessible explanation.

Feel free to share your thoughts on DeepSeek and what you think are the implications in the short-to-long term for investments. 

Hi everyone

I have just posted in "Library" the original paper of DeepSeek V3 along with a summary and some definition of key concepts to fully understand the key takeaways of the paper. You can  also find the paper from the original github source.

For those new to LLMs or just starting to explore them, this video provides a straightforward and accessible explanation.

Feel free to share your thoughts on DeepSeek and what you think are the implications in the short-to-long term for investments. 

Hi everyone

I have just posted in "Library" the original paper of DeepSeek V3 along with a summary and some definition of key concepts to fully understand the key takeaways of the paper. You can  also find the paper from the original github source.

For those new to LLMs or just starting to explore them, this video provides a straightforward and accessible explanation.

Feel free to share your thoughts on DeepSeek and what you think are the implications in the short-to-long term for investments. 

Further update on this by Jonathan Ross, the co-founder and CEO of Groq and one of the designers of the TPU original chip when working at Google/Alphabet.

video

Hi everyone

I have just posted in "Library" the original paper of DeepSeek V3 along with a summary and some definition of key concepts to fully understand the key takeaways of the paper. You can  also find the paper from the original github source.

For those new to LLMs or just starting to explore them, this video provides a straightforward and accessible explanation.

Feel free to share your thoughts on DeepSeek and what you think are the implications in the short-to-long term for investments. 

Great video

Some takeaways from the video for those with no time to watch it:

DeepSeek's Breakthrough and Market Challenges

• High-quality data distillation has been central to DeepSeek's success.
• Data privacy concerns (storage in Chinese servers) hinder adoption by global corporations.
• Pricing pressure on OpenAI: DeepSeek's lower-cost models challenge OpenAI, which may need to open access to retain retail users.
• Limited impact on basic tasks: DeepSeek's model may not significantly outperform existing solutions in simple applications like text summarisation.
• China vs. Western AI firms: Western AI companies face stricter regulations compared to China-based teams like DeepSeek.

GenAI Development and Regulation

• Blurred line between R&D and corporate theft when advancing GenAI models.
• Europe lags in AI entrepreneurship and should foster a risk-taking culture to stay competitive.
• Reliable AI for critical applications (e.g., medical diagnostics) requires low hallucination and perplexity to meet regulatory standards.

Jevon's Paradox and the AI Growth Cycle

• Jevon's Paradox: As AI becomes cheaper, demand increases, creating a feedback loop of growth.
• AI development cycle:
  • Better models → More developer demand → More inference demand → More model training → Cycle repeats.

Hardware & Infrastructure Implications

• Nvidia and AI hardware players will benefit long-term, as AI growth is still in early stages.
• Lower OPEX in AI development → Higher AI adoption → Increased demand for Nvidia's hardware.
• Inference (using models) will become a bigger business than training models.
• Stargate Project: A major bet on GenAI infrastructure to achieve dominance through scale.

MoE (Mixture of Experts) & Model Development

• MoE models will dominate due to:
  • Higher efficiency (fewer active parameters per task).
  • Scalability (flexible parameter usage).
  • Lower training costs and higher performance for specialised tasks.
• Analogy: A traditional LLM consults everyone in a room; an MoE LLM consults only the experts (e.g., CFA members for investment questions).
• Smaller, retrainable models: MoE and DeepSeek's new architecture allow smaller, easily updated models.

AI Adoption and the Future of Work

• More data ≠ better models: Quality over quantity is key in AI training.
• Training scales with user base, inference scales with usage.
• Human error tolerance varies by task:
  • Lower tolerance (e.g., car accidents, trading losses).
  • Higher tolerance (e.g., office tasks, writing errors).
  • Explains why GenAI adoption in office work is faster than in fields like autonomous driving.
• Cybersecurity & AI fuel a new Cold War as nations leverage GenAI in cyber warfare.
• Prompt engineering will become a key skill for the workforce to stay relevant.

Further update on this by Jonathan Ross, the co-founder and CEO of Groq and one of the designers of the TPU original chip when working at Google/Alphabet.

video

Hi everyone

I have just posted in "Library" the original paper of DeepSeek V3 along with a summary and some definition of key concepts to fully understand the key takeaways of the paper. You can  also find the paper from the original github source.

For those new to LLMs or just starting to explore them, this video provides a straightforward and accessible explanation.

Feel free to share your thoughts on DeepSeek and what you think are the implications in the short-to-long term for investments. 

To add to this -

Dylan Patel from SemiAnalysis has done some great deep-dive work on the total potential true traning costs end-to-end.

• total server CapEx ~$1.6B
• operating costs for clusters ~$0.9B

Todor

Great video

Some takeaways from the video for those with no time to watch it:

DeepSeek's Breakthrough and Market Challenges

• High-quality data distillation has been central to DeepSeek's success.
• Data privacy concerns (storage in Chinese servers) hinder adoption by global corporations.
• Pricing pressure on OpenAI: DeepSeek's lower-cost models challenge OpenAI, which may need to open access to retain retail users.
• Limited impact on basic tasks: DeepSeek's model may not significantly outperform existing solutions in simple applications like text summarisation.
• China vs. Western AI firms: Western AI companies face stricter regulations compared to China-based teams like DeepSeek.

GenAI Development and Regulation

• Blurred line between R&D and corporate theft when advancing GenAI models.
• Europe lags in AI entrepreneurship and should foster a risk-taking culture to stay competitive.
• Reliable AI for critical applications (e.g., medical diagnostics) requires low hallucination and perplexity to meet regulatory standards.

Jevon's Paradox and the AI Growth Cycle

• Jevon's Paradox: As AI becomes cheaper, demand increases, creating a feedback loop of growth.
• AI development cycle:
  • Better models → More developer demand → More inference demand → More model training → Cycle repeats.

Hardware & Infrastructure Implications

• Nvidia and AI hardware players will benefit long-term, as AI growth is still in early stages.
• Lower OPEX in AI development → Higher AI adoption → Increased demand for Nvidia's hardware.
• Inference (using models) will become a bigger business than training models.
• Stargate Project: A major bet on GenAI infrastructure to achieve dominance through scale.

MoE (Mixture of Experts) & Model Development

• MoE models will dominate due to:
  • Higher efficiency (fewer active parameters per task).
  • Scalability (flexible parameter usage).
  • Lower training costs and higher performance for specialised tasks.
• Analogy: A traditional LLM consults everyone in a room; an MoE LLM consults only the experts (e.g., CFA members for investment questions).
• Smaller, retrainable models: MoE and DeepSeek's new architecture allow smaller, easily updated models.

AI Adoption and the Future of Work

• More data ≠ better models: Quality over quantity is key in AI training.
• Training scales with user base, inference scales with usage.
• Human error tolerance varies by task:
  • Lower tolerance (e.g., car accidents, trading losses).
  • Higher tolerance (e.g., office tasks, writing errors).
  • Explains why GenAI adoption in office work is faster than in fields like autonomous driving.
• Cybersecurity & AI fuel a new Cold War as nations leverage GenAI in cyber warfare.
• Prompt engineering will become a key skill for the workforce to stay relevant.

Further update on this by Jonathan Ross, the co-founder and CEO of Groq and one of the designers of the TPU original chip when working at Google/Alphabet.

video

Hi everyone

I have just posted in "Library" the original paper of DeepSeek V3 along with a summary and some definition of key concepts to fully understand the key takeaways of the paper. You can  also find the paper from the original github source.

For those new to LLMs or just starting to explore them, this video provides a straightforward and accessible explanation.

Feel free to share your thoughts on DeepSeek and what you think are the implications in the short-to-long term for investments. 

Great video

Some takeaways from the video for those with no time to watch it:

DeepSeek's Breakthrough and Market Challenges

• High-quality data distillation has been central to DeepSeek's success.
• Data privacy concerns (storage in Chinese servers) hinder adoption by global corporations.
• Pricing pressure on OpenAI: DeepSeek's lower-cost models challenge OpenAI, which may need to open access to retain retail users.
• Limited impact on basic tasks: DeepSeek's model may not significantly outperform existing solutions in simple applications like text summarisation.
• China vs. Western AI firms: Western AI companies face stricter regulations compared to China-based teams like DeepSeek.

GenAI Development and Regulation

• Blurred line between R&D and corporate theft when advancing GenAI models.
• Europe lags in AI entrepreneurship and should foster a risk-taking culture to stay competitive.
• Reliable AI for critical applications (e.g., medical diagnostics) requires low hallucination and perplexity to meet regulatory standards.

Jevon's Paradox and the AI Growth Cycle

• Jevon's Paradox: As AI becomes cheaper, demand increases, creating a feedback loop of growth.
• AI development cycle:
  • Better models → More developer demand → More inference demand → More model training → Cycle repeats.

Hardware & Infrastructure Implications

• Nvidia and AI hardware players will benefit long-term, as AI growth is still in early stages.
• Lower OPEX in AI development → Higher AI adoption → Increased demand for Nvidia's hardware.
• Inference (using models) will become a bigger business than training models.
• Stargate Project: A major bet on GenAI infrastructure to achieve dominance through scale.

MoE (Mixture of Experts) & Model Development

• MoE models will dominate due to:
  • Higher efficiency (fewer active parameters per task).
  • Scalability (flexible parameter usage).
  • Lower training costs and higher performance for specialised tasks.
• Analogy: A traditional LLM consults everyone in a room; an MoE LLM consults only the experts (e.g., CFA members for investment questions).
• Smaller, retrainable models: MoE and DeepSeek's new architecture allow smaller, easily updated models.

AI Adoption and the Future of Work

• More data ≠ better models: Quality over quantity is key in AI training.
• Training scales with user base, inference scales with usage.
• Human error tolerance varies by task:
  • Lower tolerance (e.g., car accidents, trading losses).
  • Higher tolerance (e.g., office tasks, writing errors).
  • Explains why GenAI adoption in office work is faster than in fields like autonomous driving.
• Cybersecurity & AI fuel a new Cold War as nations leverage GenAI in cyber warfare.
• Prompt engineering will become a key skill for the workforce to stay relevant.

Further update on this by Jonathan Ross, the co-founder and CEO of Groq and one of the designers of the TPU original chip when working at Google/Alphabet.

video

Hi everyone

I have just posted in "Library" the original paper of DeepSeek V3 along with a summary and some definition of key concepts to fully understand the key takeaways of the paper. You can  also find the paper from the original github source.

For those new to LLMs or just starting to explore them, this video provides a straightforward and accessible explanation.

Feel free to share your thoughts on DeepSeek and what you think are the implications in the short-to-long term for investments. 

Another follow up analysis from Jonathan Ross on the future of Training vs. Inference.

video

Great video

Some takeaways from the video for those with no time to watch it:

DeepSeek's Breakthrough and Market Challenges

• High-quality data distillation has been central to DeepSeek's success.
• Data privacy concerns (storage in Chinese servers) hinder adoption by global corporations.
• Pricing pressure on OpenAI: DeepSeek's lower-cost models challenge OpenAI, which may need to open access to retain retail users.
• Limited impact on basic tasks: DeepSeek's model may not significantly outperform existing solutions in simple applications like text summarisation.
• China vs. Western AI firms: Western AI companies face stricter regulations compared to China-based teams like DeepSeek.

GenAI Development and Regulation

• Blurred line between R&D and corporate theft when advancing GenAI models.
• Europe lags in AI entrepreneurship and should foster a risk-taking culture to stay competitive.
• Reliable AI for critical applications (e.g., medical diagnostics) requires low hallucination and perplexity to meet regulatory standards.

Jevon's Paradox and the AI Growth Cycle

• Jevon's Paradox: As AI becomes cheaper, demand increases, creating a feedback loop of growth.
• AI development cycle:
  • Better models → More developer demand → More inference demand → More model training → Cycle repeats.

Hardware & Infrastructure Implications

• Nvidia and AI hardware players will benefit long-term, as AI growth is still in early stages.
• Lower OPEX in AI development → Higher AI adoption → Increased demand for Nvidia's hardware.
• Inference (using models) will become a bigger business than training models.
• Stargate Project: A major bet on GenAI infrastructure to achieve dominance through scale.

MoE (Mixture of Experts) & Model Development

• MoE models will dominate due to:
  • Higher efficiency (fewer active parameters per task).
  • Scalability (flexible parameter usage).
  • Lower training costs and higher performance for specialised tasks.
• Analogy: A traditional LLM consults everyone in a room; an MoE LLM consults only the experts (e.g., CFA members for investment questions).
• Smaller, retrainable models: MoE and DeepSeek's new architecture allow smaller, easily updated models.

AI Adoption and the Future of Work

• More data ≠ better models: Quality over quantity is key in AI training.
• Training scales with user base, inference scales with usage.
• Human error tolerance varies by task:
  • Lower tolerance (e.g., car accidents, trading losses).
  • Higher tolerance (e.g., office tasks, writing errors).
  • Explains why GenAI adoption in office work is faster than in fields like autonomous driving.
• Cybersecurity & AI fuel a new Cold War as nations leverage GenAI in cyber warfare.
• Prompt engineering will become a key skill for the workforce to stay relevant.

Further update on this by Jonathan Ross, the co-founder and CEO of Groq and one of the designers of the TPU original chip when working at Google/Alphabet.

video

Hi everyone

I have just posted in "Library" the original paper of DeepSeek V3 along with a summary and some definition of key concepts to fully understand the key takeaways of the paper. You can  also find the paper from the original github source.

For those new to LLMs or just starting to explore them, this video provides a straightforward and accessible explanation.

Feel free to share your thoughts on DeepSeek and what you think are the implications in the short-to-long term for investments. 

More new developments on LLMs coming from  Europe:

• Mistral on Tuesday launched Europe's first AI reasoning model, which uses logical thinking to create a response, as it tries to keep pace with American and Chinese rivals at the forefront of AI development.
• The French startup has attempted to differentiate itself by championing its European roots, winning the support of French President Emmanuel Macron, as well as making some of its models open source in contrast to the proprietary offerings of OpenAI or Alphabet's Google.
• Mistral is considered Europe's best shot at having a home-grown AI competitor, but has lagged behind in terms of market share and revenue.
• Reasoning models use chain-of-thought techniques - a process that generates answers with intermediate reasoning abilities when solving complex problems.
• They could also be a promising path forward in advancing AI's capabilities as the traditional approach of building ever-bigger large language models by adding more data and computing power begins to hit limitations.
• For Mistral, which was valued by venture capitalists at $6.2 billion, an industry shift away from "scaling up" could give it a window to catch up against better capitalized rivals.
• China's DeepSeek broke through as a viable competitor in January through its low-cost, open-sourced AI models, including one for reasoning.
• OpenAI was the first to launch its reasoning models last year, followed by Google a few months later.
• Meta, which also offers its models open-sourced, has not yet released a standalone reasoning model, though it said its latest top-shelf model has reasoning capabilities.
• Mistral is launching an open-sourced Magistral Small model and a more powerful version called Magistral Medium for business customers.
• "The best human thinking isn't linear - it weaves through logic, insight, uncertainty, and discovery. Reasoning language models have enabled us to augment and delegate complex thinking and deep understanding to AI," Mistral said. 
• American companies have mostly kept their most advanced models proprietary, though a handful, such as Meta, has released open-source models. In contrast, Chinese firms ranging from DeepSeek to Alibaba have taken the open-source path to demonstrate their technological capabilities.
• Mistral Small is available for download on Hugging Face's platform and can reason in languages including English, French, Spanish, Arabic and simplified Chinese.
  
  

sources: 

• https://www.axios.com/2025/06/10/mistral-ai-reasoning-models-open-source
• https://www.reuters.com/business/frances-mistral-launches-europes-first-ai-reasoning-model-2025-06-10/

Feel free to add information or comment on this new LLM development

Another follow up analysis from Jonathan Ross on the future of Training vs. Inference.

video

Great video

Some takeaways from the video for those with no time to watch it:

DeepSeek's Breakthrough and Market Challenges

• High-quality data distillation has been central to DeepSeek's success.
• Data privacy concerns (storage in Chinese servers) hinder adoption by global corporations.
• Pricing pressure on OpenAI: DeepSeek's lower-cost models challenge OpenAI, which may need to open access to retain retail users.
• Limited impact on basic tasks: DeepSeek's model may not significantly outperform existing solutions in simple applications like text summarisation.
• China vs. Western AI firms: Western AI companies face stricter regulations compared to China-based teams like DeepSeek.

GenAI Development and Regulation

• Blurred line between R&D and corporate theft when advancing GenAI models.
• Europe lags in AI entrepreneurship and should foster a risk-taking culture to stay competitive.
• Reliable AI for critical applications (e.g., medical diagnostics) requires low hallucination and perplexity to meet regulatory standards.

Jevon's Paradox and the AI Growth Cycle

• Jevon's Paradox: As AI becomes cheaper, demand increases, creating a feedback loop of growth.
• AI development cycle:
  • Better models → More developer demand → More inference demand → More model training → Cycle repeats.

Hardware & Infrastructure Implications

• Nvidia and AI hardware players will benefit long-term, as AI growth is still in early stages.
• Lower OPEX in AI development → Higher AI adoption → Increased demand for Nvidia's hardware.
• Inference (using models) will become a bigger business than training models.
• Stargate Project: A major bet on GenAI infrastructure to achieve dominance through scale.

MoE (Mixture of Experts) & Model Development

• MoE models will dominate due to:
  • Higher efficiency (fewer active parameters per task).
  • Scalability (flexible parameter usage).
  • Lower training costs and higher performance for specialised tasks.
• Analogy: A traditional LLM consults everyone in a room; an MoE LLM consults only the experts (e.g., CFA members for investment questions).
• Smaller, retrainable models: MoE and DeepSeek's new architecture allow smaller, easily updated models.

AI Adoption and the Future of Work

• More data ≠ better models: Quality over quantity is key in AI training.
• Training scales with user base, inference scales with usage.
• Human error tolerance varies by task:
  • Lower tolerance (e.g., car accidents, trading losses).
  • Higher tolerance (e.g., office tasks, writing errors).
  • Explains why GenAI adoption in office work is faster than in fields like autonomous driving.
• Cybersecurity & AI fuel a new Cold War as nations leverage GenAI in cyber warfare.
• Prompt engineering will become a key skill for the workforce to stay relevant.

Further update on this by Jonathan Ross, the co-founder and CEO of Groq and one of the designers of the TPU original chip when working at Google/Alphabet.

video

Hi everyone

I have just posted in "Library" the original paper of DeepSeek V3 along with a summary and some definition of key concepts to fully understand the key takeaways of the paper. You can  also find the paper from the original github source.

For those new to LLMs or just starting to explore them, this video provides a straightforward and accessible explanation.

Feel free to share your thoughts on DeepSeek and what you think are the implications in the short-to-long term for investments. 

Hi everyone

I have just posted in "Library" the original paper of DeepSeek V3 along with a summary and some definition of key concepts to fully understand the key takeaways of the paper. You can  also find the paper from the original github source.

For those new to LLMs or just starting to explore them, this video provides a straightforward and accessible explanation.

Feel free to share your thoughts on DeepSeek and what you think are the implications in the short-to-long term for investments. 

Meet GPT-5

GPT-5 is OpenAI's newest flagship model. It's a big step up in reasoning, speed, and "do-the-task" reliability, and it's built to work across text, images, audio, and video.

What it's good at

• Stronger reasoning & planning: tackles multi-step problems with fewer mistakes and better uncertainty handling. 

• Agentic skills: pairs with ChatGPT's new agent to operate software for you (think: "book this flight and file the expense") while staying within clear permissions. 

• Multimodal I/O: understand and generate across text, images, and speech, and reason over mixed inputs. 

• Long context: far longer working memory than prior models, enabling large docs, datasets, and end-to-end workflows.

• Developer-friendly: benefits from recent API advances like structured outputs and the newer 4.x family groundwork. 

Where you can use it

• ChatGPT (personal and work plans), with GPT-5 as the default "just works" model.

• OpenAI API for apps/automation, alongside mini/nano variants for cost/latency trade-offs.

Why it matters

Compared to earlier GPT-4.x models, GPT-5 is more reliable at complex tasks, better at following instructions, and faster-so it handles end-to-end jobs (analysis → decision → action) with less hand-holding. 

https://openai.com/index/introducing-gpt-5/

Hi everyone

I have just posted in "Library" the original paper of DeepSeek V3 along with a summary and some definition of key concepts to fully understand the key takeaways of the paper. You can  also find the paper from the original github source.

For those new to LLMs or just starting to explore them, this video provides a straightforward and accessible explanation.

Feel free to share your thoughts on DeepSeek and what you think are the implications in the short-to-long term for investments.