Chronicles of 2025

[user137]

Bought some ASML stock and thinking about topping up TSM as well.
Buy the shovel factory’s stock in the gold rush…

[PhilMongoose]

I regret selling both ASML and TSM!

[Mirager]

@PhilMongoose

It’s a strange reaction. I think they are interpreting this wrongly. While Deepseeks shows that you can train more efficiently, I don’t think this implies fewer GPU sales. If anything, the move to test time compute implies a lot more computing power will be needed.

Don’t try to be reasonable with the toddler (the market). Until we have a nice big correction (so people’s brains go back to their unagitated but still stupid selves) we’ll have swings every time something AI farts the wrong way.

Sticking to my plan, saving cash and making no movements until June 2025.

[user137]

@Mirager

making no movements until June 2025

what’s happening in June 2025?

[Mirager]

It’d have been a year of saving and not buying and building up cash, set up as a plan 8 months ago. Just expecting a lot of volatility this year, so if we get gains, great, of we get drops, great as I’ll have liquidity.

Not moving a hair for anything less than 10-15% drop or more.

[ma0]

@Mirager

10-15% drop or more.

From ATH or daily/weekly ?

[almi]

@almi

Results
All results here are only about % return. Absolute return varies due to different amount invested.
Note, for baseline, the return of investment in % would be the same if I invest 1000$ or e.g. 2000$/month (corresponding to 240 buy-events, but 2 on the same day; disregarding fees).

• Baseline - Average 10-y return: 60.7% Std: 14.9% Buy-events (average): 120
• Immediate Buy-the-Dip - Average 10-y return: 63.5% Std: 15.8% Buy-events (average): 160
• Short-Delayed Buy-the-Dip - Average 10-y return: 63.5% Std: 16.2% Buy-events (average): 175
• Cool-Off Buy-the-Dip - Average 10-y return: 62.2% Std: 15.3% Buy-events (average): 131
• Baseline 50-50 Dip - Average 10-y return: 60.7% Std: 14.9% Buy-events (average): 160
• Hardcore Buy-the-Dip - Average 10-y return: 60.6% Std: 14.9% Buy-events (average): 120

We haven’t yet reached the -2% dip used in the calculations above, but if we do, it’s at least historically not disadvantageous to play some immediate buy-the-dip market timing*.

(*With the money that should in theory already be invested ).

[Abs_max]

I don’t think it’s strange
It simple is a reaction that to win in AI race , you cannot assume Mag 7 will continue their domination or Monopoly. Thus their high valuations are being challenged. It also questions the growth estimates built in for the Mag 7 for coming years

It’s not about NVDIA being best or not. It’s about the fact that Deepseek was able to build these models without the most advanced (and highest price) chips from NVDA.

I think the companies who are mainly the users of tech and not producers of AI should ideally benefit because they might be able to become more productive at lower cost

[Mirager]

I always go from ATH, caveat that the dip may never come.

[user137]

@Abs_max

It’s not about NVDIA being best or not. It’s about the fact that Deepseek was able to build these models without the most advanced (and highest price) chips from NVDA.

I think the companies who are mainly the users of tech and not producers of AI should ideally benefit because they might be able to become more productive at lower cost

I’m very skeptical of the fact that the Mag7 engineers would consistently overlook options how to do a better model with not enough and outdated hardware and still be 2 orders of magnitude cheaper on the cost base… how is this even possible?

[Mirager]

Because this is not where their competitive advantage is, in my opinion.

The MAG7 are probably the best funded companies, with the best brains they can get from the whole globe, why would they look backward?

[Your_Full_Name]

@user137

I’m very skeptical of the fact that the Mag7 engineers would consistently overlook options

I’m not.

When you have too much cash sitting on your balance sheet you tend to no longer be scrappy and just throw money at the problem at hand … I have seen this myself in my last couple of years at Hoolie.

[Your_Full_Name]

I like this description of what’s happening today:

(Source)

[Abs_max]

Necessity is mother of invention

[user137]

@Your_Full_Name

When you have too much cash sitting on your balance sheet you tend to no longer be scrappy and just throw money at the problem at hand … I have seen this myself in my last couple of years at Hoolie.

OpenAI is flush with cash BUT they are massively unprofitable, so even though they can run with the cash runway, they can’t hide the fact that they need to be profitable at some point.

As the TAM of AI is exorbitant, all market players have very good incentives to work on effective downscaling. (do you hear that, Project Stargate? )

I would believe the Chinese did something comparable on somewhat older devices, but what I’m hearing is that it’s massively smaller scale that is massively better and is made by solely inland Chinese without any western influence or returning Asians from the Valley, which to me sounds like pure play propaganda.

[nabalzbhf]

@user137

I would believe the Chinese did something comparable on somewhat older devices, but what I’m hearing is that it’s massively smaller scale that is massively better and is made by solely inland Chinese without any western influence or returning Asians from the Valley, which to me sounds like pure play propaganda.

If you want western reporting on the situation: Subscribe to read

[almi]

Interesting article. But in the end we still have to trust that what they claim they used for training is what they actually used. The number of chips, training time, education origin of their staff etc. could all be wrong and it would be hard to prove that. Probably also irrelevant for most people that are just interested in the final model.

The good thing is that the model itself and the documentation is open-source, so whoever and with whatever resources built it, did a good job.

[nabalzbhf]

@almi

The good thing is that the model itself and the documentation is open-source, so whoever and with whatever resources built it, did a good job.

Yeah there’s a paper, should be easy to verify the claims (I think huggingface among others are trying to replicate it)

[Your_Full_Name]

@user137

what I’m hearing is that it’s massively smaller scale that is massively better and is made by solely inland Chinese without any western influence or returning Asians from the Valley, which to me sounds like pure play propaganda.

The market seems to disagree, though, and we all know the market is always efficient.

I’m actually a little skeptical myself (along the lines of your thoughts expressed), but then again that’s always the doubt one might have when a new disruptor enters the arena.
We’ll find out eventually.

(Meanwhile, my Broadcom position is down about 16% 20% )

[TeaGhost]

We don‘t know yet if this was real or political warfare by the chinese. Lets wait for a month.

One thought - maybe chinese is a better starting point to train a LLM - more characters…