DeepReflection, Private Credit, Local AI

DeepSeek Reflection

Straight to the point: competition is a good thing. This DeepSeek episode has re-established a certain level of confidence that we can improve AI models through algorithmic breakthroughs. We don’t have to rely only on astronomically increasing budgets, like $500B for Stargate.

Knowledgeable people have said the model did not cost $6m and was not trained on old chips. China can’t buy NVDA chips directly, but that’s relatively easy to circumvent. It’s also ironic that as we’re working on banning TikTok, Red Note rises to the top of the app Charts, and then DeepSeek hits number 1. US IPhone users can’t get enough Chinese AI apps. It’s a waste of time to ban each company individually.

DeepSeek is an open weight model. The model is freely available to download, run and tweak without paying a subscription. The model is not considered truly open source because we do not have access to the data the model has been trained on.

The AI tsunami is not stopping for a single product release. Full speed ahead.

Private Credit Comments

Our position has been for some time that there are only two areas of the fixed income markets worth investing in: treasuries and private credit. This is not specific advice, every person and circumstance is unique. Treasuries yield just over 4% currently as a cash equivalent and/or small bond portfolio. Private credit yields 9% at the lower end, 10-12% in the middle and up to 15% at the higher end.

Private credit is just a collection of not-publicly-traded debt. The yields are much higher at a macro level for two reasons: higher federal interest rates and the collapse of regional banking (like SVB and First Republic). The yields are also higher because there are many scenarios where a borrower is willing to pay 15%+ on a loan if it’s a short duration, they are earning a higher return on capital on something else, or both.

People should be extra careful when investing in private credit through major brokerages like UBS or Goldman. Large investment firms have riddled their core menu of private credit funds with ridiculously high and fully unnecessary extra fees.

Please contact us if you suspect this and we’ll show you where to look.

Local Private AI

If you search “DeepSeek Raspberry Pi” you’ll get a peak at what’s to come for private at home LLMs. All major AI tools today are accessed via the internet. If you’re a devoted hobbyist or just interested in this type of thing, you have already read about downloading the NVDA LLM model and running it locally on your gaming computer. That’s still a relatively expensive endeavor that as of last year would cost around $2-5k.

You can now download a copy of DeepSeek (it’s open weight) and run it on a Raspberry Pi for about $50 total. It will be slow, but it will work. You can make it faster by adding on a GPU to process the LLM calculations for between $250-$1000k.

The problem with internet-accessed LLMs is there is no privacy. Most people are too busy and live in a world that’s moving too fast to care about privacy, but it is a factor. Realistically it would be nice to have an at-home not-connected-to-the-internet LLM to run sensitive data or just be better aligned with data privacy. Technically that now costs just $50 as a one time payment, not a subscription. That price will not fall lower and it doesn’t need to, but the models will get significantly better.

It’s important to note that NVDA already released a small computer called the Jetson Orin Nano that essentially does this for $250 total. It’s a niche performance product that’s designed to add AI in a small place, like on a drone, for example.