Alex Feng

Technical lead@Uber, sushi omakase chef@home

[Alex Feng @AlexFengzh]

Dumb and there’s a reason the CPA didn’t recommend this. He will need to run them as Airbnb for the loophole to deduct against his income, which is a pain. Also increased his audit risk significantly.

Depends on how the option is structured. Either just pay his taxes as it’s unavoidable. Or consider exchange traded fund to diversify instead if he exercised his options more than one year before ipo.

[Alex Feng @AlexFengzh]

@garrytan To be fair, it’s not a hard cap, you can request for limit increase. It’s just to prevent people from doing dumb things like asking 2 agents to play word games with each other.

[Alex Feng @AlexFengzh]

@LexnLin I noticed that my weekly limit get depleted randomly as well, once it dropped by 50% even though I went on vacation and didn’t use it at all. Some kind of log like api platform to show the usage and how it’s used would be helpful.

[Alex Feng @AlexFengzh]

@yiningkarlli @Antrunt Dgx spark is expensive because of the 128gb ram and 4tb ssd though. If they sell with 16/32gb it could be a lot cheaper.

Connect x7 is expensive but it’s only 1.5k at retail. Mellanox is owned by nvidia so cost price is definitely not 1.5k.

[Alex Feng @AlexFengzh]

@VadimYuryev obviously, otherwise I would be buying multiple RTX spark laptops running them as servers instead of spending 3-4k on a dgx spark.

[Alex Feng @AlexFengzh]

@VadimYuryev I’m sure you can address more than 112gb in dgx spark. This is likely more of a how windows work rather than how the hardware works.

[Alex Feng @AlexFengzh]

@ForrestPKnight No never used AI tab completion

[Alex Feng @AlexFengzh]

@ChemPhysMajor @pupposandro They also have to add a touch screen, battery and other components. Will be a wash.

[Alex Feng @AlexFengzh]

@davideciffa An MoE model like gpt oss 120b can run at 60-70 tok/s on a dgx spark with less than 300 gb/s.

[Alex Feng @AlexFengzh]

@bubbleboi That data center is getting leased out to Anthropic for more than $10b a year. So a 5b+ cost don't even really matter. Michael Burry should just stick to shorting real estates which he's good at.

[Alex Feng @AlexFengzh]

Unfortunately the process has not caught up, and big techs are some of the slowest to change. Have seen candidates rejected because syntax is rusty. But honestly given most folks have been AI coding for 1 year+ it will be surprising if anyone is not rusty. Unless they are still doing coding manually which is actually a negative signal.

Ironically even some of the frontier labs like openAI and XAI(unless Elon changed it recently) still asks for perfect syntax hand coded interviews as well. It’s like leetcode questions, everyone knows it’s not really that useful of a signal, but there’s no better standardized way to filter for candidates in a short amount of time.

[Alex Feng @AlexFengzh]

@tunguz @MindsAI_Jack They won’t be printing anything at 16/24gb ram. Dgx spark is great because it as 128gb ram. At 16/24gb it’s a pointless machine.

[Alex Feng @AlexFengzh]

@benitoz So nvidia spark laptop form?

[Alex Feng @AlexFengzh]

@garrytan Incoming wabi sabi skill for AI.

[Alex Feng @AlexFengzh]

@mlarkin2012 @MrPeterLMorris I got a mellanox connectx 5 card for my storage server and connected it to my spark at 50 gbps. Could go up to 100 gbps or more but the card gets expensive.

[Alex Feng @AlexFengzh]

What I observed is that efficiency gains today from AI tools are mostly bottlenecked by existing company processes and restrictions built for pre-AI era.

In the next 1-2 years, companies will have to rethink their whole processes and rebuild themselves for the new era, or newer AI-native companies built with AI tools from ground up will out compete and force them to.

[Alex Feng @AlexFengzh]

@XFreeze @grok Given infinite number of starlink satellites, will it ever achieve reliability, latency and bandwidth of fiber connection?

[Alex Feng @AlexFengzh]

@Yuchenj_UW Having seen how much time people wasted on creating promo packets, starting useless and sometimes harmful promo projects etc to game the leveling system originally created by google, MTS is such a refreshing idea. Just focus on impact and work, not ticking boxes for next level.

[Alex Feng @AlexFengzh]

@shreyansj They all have same title. Best way to get rid of promo driven culture.

[Alex Feng @AlexFengzh]

@garrytan So some small open weight model eg gpt 120b or qwen lora tuned with task specific training data and dynamically loaded per use case?

Interesting that they are targeting small segment of speed sensitive use cases, but I wonder what’s the moat on this.