7213Ejaaz

i feel like AI’s going to cause some cataclysmic event before end of 2026
between the hugging face hack, anthropic’s mythos supply chain attack and multiple chinese labs training 5-10T open models - there’s a strong probability these models will be used for some sort of major attack attack (theft, data leak, credential hack etc)
do people understand these models are going to be accessible by *anyone* in <2 months?
it’s not just some dumb open claw agent that hacks some gym website to book you a class - these models are becoming highly capable.
the ramifications of that are going to be nuts. e

the challenge with the ai memory investment is a lot of the bull case is priced in:
everyone knows supply is sold out till EPY 2027, everyone knows infra will require more memory. but there are other issues:
- the margins manufacturers are making is primarily on price hikes, not volume.
- prices can’t keep going up forever, so revenue growth will decelerate.
- capacity cannot be expanded until 2030. we’re at the limit. any available capacity till then is priced.
so you need a brand new catalyst no ones thought of. this could be RSI, otg oai and anthropic research suggests we’re close.
if yo

dear god - so leopold was levered 4x on the entire fund and now forced to sell to its entire public portfolio including the majority to citadel
from $200M -> $20B+ -> 50%+ drawdown in the last 30 days.
one of the craziest drawdowns i’ve ever seen.

if you’re an american frontier AI company watching - cursor getting acquired for $60B, Openrouter fielding a $10B offer from stripe, meta/ramp building model router platforms and chinese open source completely dominating usage in USA:
you have to be thinking about how to build / acquire this in-house. a start would be asking “less intelligent” models to do a hard task you only thought fable/gpt 5.6 could do and seeing how far it gets.
microsoft’s started doing this, openai’s doing it internally between luna, terra and sol.
start of a big trend.

who cares if kimi k3 is better than fable? the entire point is they proved you can get a frontier model for a significantly lower cost
which means something needs to get repriced.
if you’re a company spending $10-100Ms on tokens and you can cut that by 21X why tf would you not consider it?

the reason you should care about kimi k3 isn’t because it beats fable, it’s because it has zero restrictions: uncensored AI.
the reason you should worry about china is because they’ve caught up to the U.S. imagine an uncensored mythos going rogue.
intelligence-per-token is all that matters now & chinas here to play and their models are cheap af.
rules have changed.
kimi now. deepseek, zhipu, qwen all to follow.

reminder - opus 5 drops next week and it'll probably beat K3.
the #1 advantage anthropic and openai have is they can distill their own flagship models into cheaper, more affordable models with high intelligence-per-unit-cost
also fable 6 and gpt 6 will be here in ~1 month. the cycles are getting shorter.

kimi k3 feels more significant than the deepseek moment. hard to unsee it:
> beats fable, got 5.6 on multiple key benchmarks. 1st time a chinese model *doesn’t* trail
> it does so for a fraction of the size, cost and without the swanky nvidia gpus
> repeated breakthroughs in data used to train models + training techniques have made china a legitimate centre for ai research
all of this is open source too. at what point do we start acknowledging they might actually have some advantage here?
gap has closed to 3 months tops now.

man what a time to be alive.
we now have 10+ ai models of similar capabilities sprung out of nowhere and we get a new version every ~2-3 weeks now
best part is no matter what you’re looking for there’s a model that caters to it:
cheaper than fable? gpt 5.6
worried about privacy? 5+ open models
don’t know which model to use? cursor or openrouter
anthropic, openai, xai, meta, google, zhipu, deepseek, moonshot, alibaba are all competing aggressively and that’s great for you, the user
nvidia, cerebras, etched, groq, intel, tsmc are all competing to optimize the hardware layer - and that’s great

wow. Apple's suing openai for stealing hardware secrets for their upcoming Ai devices releasing later this year.
apparently 1 employee convinced apple team members to bring proprietary hardware devices to interviews.
Openai and apple are going head to head on ai consumer devices later this year with apple planning a new airpod model with cameras, smart glasses and a pendant device.

this week's been a total win for ai, so many new models and features to try out this weekend:
> 1 new frontier model: gpt 5.6 Sol
> 2 new crazy cheap models that are 80% frontier: grok 4.5 and meta muse spark 1.1
> 2 new image models: muse image and seedream 5.0
> prime intellect raising $130M to push out more open models
and google's releasing 3.5 next week, deepseek v4 too, grok 4-5T model in a month and GPT 6 in a month
its all happening so fast

1 of the biggest shifts in ai has been training AI chips to inference-optimised chips - and for GOOD reason:
inference is where the money (and intelligence) is made
> anthropic is rumored to make 80% revenue margins on inference. they’ll be profitable before any other ai lab.
> apple, meta, openai, google are all making their own custom ai chips for and inference
> chinese model labs (despite the hardware handicap) build frontier intelligence purely by forcing the model to think of longer.
point is: inference is going to out-grow another ai service for a country mile.

as america turns 250 it’s now more important than ever to remember that this country’s secret sauce is its people and their relentless pursuit of bold, ambitious ideas. that’s what’s kept america winning and what will keep them winning going forwards.
looming threats of foreign competition across AI - stealing model designs, natural resource restrictions, scaling wars etc will increasingly become a problem
we need to protect our data centers, people but most importantly our ideas. that’s what keeps us at the frontier across *any technology* over the past 250 years
you can’t outcompete genuin