Offshore
Photo
memenodes
Megan Fox was the world’s first Hot Girl
We didn’t know girls could be that hot https://t.co/eAxAGSYDWQ
tweet
Megan Fox was the world’s first Hot Girl
We didn’t know girls could be that hot https://t.co/eAxAGSYDWQ
tweet
Offshore
Photo
memenodes
Your girl and her co-worker whenever yall get into an argument https://t.co/qbzCxECeqa
tweet
Your girl and her co-worker whenever yall get into an argument https://t.co/qbzCxECeqa
tweet
Offshore
Video
memenodes
my girl checking on me after busy all day
tweet
my girl checking on me after busy all day
Spider-Man (2002) https://t.co/oVGkn9s9YH tweet
Offshore
Video
God of Prompt
RT @godofprompt: AI agents can now buy services and products on Contra.
If you’re too busy to search through millions of freelancers, send OpenClaw.
This is a massive advancement.
tweet
RT @godofprompt: AI agents can now buy services and products on Contra.
If you’re too busy to search through millions of freelancers, send OpenClaw.
This is a massive advancement.
Introducing Contra Payments.
The first payments platform that lets you sell to AI Agents.
RT + Comment “Contra” and I’ll send you 100 products AI agents are looking for. https://t.co/XgPgfq4pRW tweet
Offshore
Video
memenodes
You start earning and finally understand why dad never spent on himself
https://t.co/XEhfqj8jC9
tweet
You start earning and finally understand why dad never spent on himself
https://t.co/XEhfqj8jC9
tweet
Offshore
Video
memenodes
Girls will look at you like this and expect you to know everything https://t.co/OyAJVU3sYJ
tweet
Girls will look at you like this and expect you to know everything https://t.co/OyAJVU3sYJ
tweet
Offshore
Video
memenodes
When I get a notification that somebody has liked a post I made two years ago https://t.co/RoxY9Q62Ze
tweet
When I get a notification that somebody has liked a post I made two years ago https://t.co/RoxY9Q62Ze
tweet
Michael Fritzell (Asian Century Stocks)
4% of GDP is spent on industrial subsidies. Cutting subsidies would presumably affect corporate competitiveness
tweet
4% of GDP is spent on industrial subsidies. Cutting subsidies would presumably affect corporate competitiveness
IMF calls on China to cut industrial subsidies in half https://t.co/1E4sUzlmyr tweet
X (formerly Twitter)
Financial Times (@FT) on X
IMF calls on China to cut industrial subsidies in half https://t.co/1E4sUzlmyr
Offshore
Video
Michael Fritzell (Asian Century Stocks)
RT @RocksOver: This one got taken out by Bain at a 117% premium. I had already sold my small position as I’d never quite gotten comfortable with the business model. But Inforich was trading dirt cheap for the growth profile so not a surprise that PE got interested. It’s a target rich environment in Japan right now.
tweet
RT @RocksOver: This one got taken out by Bain at a 117% premium. I had already sold my small position as I’d never quite gotten comfortable with the business model. But Inforich was trading dirt cheap for the growth profile so not a surprise that PE got interested. It’s a target rich environment in Japan right now.
Inforich ($9338.T) is an interesting stock I’ve never seen discussed on here. Leading provider of mobile phone charging services in Japan, with 10x scale of nearest competitor. Shares are down 40% YTD and now trade at around 10.5x FY25 guidance while the company is in hyper growth mode. expected to post around 50% revenue and EPS growth in FY25. tweet
X (formerly Twitter)
Turning over rocks (@RocksOver) on X
Inforich ($9338.T) is an interesting stock I’ve never seen discussed on here. Leading provider of mobile phone charging services in Japan, with 10x scale of nearest competitor. Shares are down 40% YTD and now trade at around 10.5x FY25 guidance while the…
Offshore
Photo
God of Prompt
RT @rryssf_: the same Google Research team that invented speculative decoding just dropped a paper arguing you should send your prompt twice 🤯
the idea sounds absurd until you understand why it works.
llms process tokens left-to-right. each token can only attend to what came before it. so when you write a long prompt with context first and a question at the end, the context tokens were processed blind to the question. they had no idea what was being asked.
prompt repetition fixes this by simply duplicating the input: <query<query.
now every token in the second copy can attend to every token in the first. the question sees the context. the context sees the question. the asymmetry disappears.
the results across 7 models (Gemini, GPT-4o, Claude, DeepSeek) and 7 benchmarks:
> 47 wins out of 70 tests, with 0 losses
> one model jumped from 21% to 97% on a name-lookup task
> no increase in output length
> no meaningful increase in latency
that last point is the clever part. doubling the input only affects the prefill stage, which runs in parallel on modern GPUs. the slow part of inference is token-by-token generation, and that doesn't change at all.
the catch: when reasoning is enabled ("think step by step"), the gains mostly vanish. 5 wins, 1 loss, 22 ties. the authors think reasoning models already repeat the prompt internally as part of their chain-of-thought. so you're just doing what the model was going to do anyway.
what's interesting is who wrote this. Yaniv Leviathan and Matan Kalman are the researchers behind speculative decoding, which became standard infrastructure across the industry for faster inference. they clearly think about the gap between how transformers actually work and how we use them.
the real implication isn't "copy-paste your prompts." it's that causal attention creates a structural blind spot we've been ignoring. the first tokens in your prompt are always the least informed.
and for non-reasoning tasks, a second pass is the cheapest fix available.
tweet
RT @rryssf_: the same Google Research team that invented speculative decoding just dropped a paper arguing you should send your prompt twice 🤯
the idea sounds absurd until you understand why it works.
llms process tokens left-to-right. each token can only attend to what came before it. so when you write a long prompt with context first and a question at the end, the context tokens were processed blind to the question. they had no idea what was being asked.
prompt repetition fixes this by simply duplicating the input: <query<query.
now every token in the second copy can attend to every token in the first. the question sees the context. the context sees the question. the asymmetry disappears.
the results across 7 models (Gemini, GPT-4o, Claude, DeepSeek) and 7 benchmarks:
> 47 wins out of 70 tests, with 0 losses
> one model jumped from 21% to 97% on a name-lookup task
> no increase in output length
> no meaningful increase in latency
that last point is the clever part. doubling the input only affects the prefill stage, which runs in parallel on modern GPUs. the slow part of inference is token-by-token generation, and that doesn't change at all.
the catch: when reasoning is enabled ("think step by step"), the gains mostly vanish. 5 wins, 1 loss, 22 ties. the authors think reasoning models already repeat the prompt internally as part of their chain-of-thought. so you're just doing what the model was going to do anyway.
what's interesting is who wrote this. Yaniv Leviathan and Matan Kalman are the researchers behind speculative decoding, which became standard infrastructure across the industry for faster inference. they clearly think about the gap between how transformers actually work and how we use them.
the real implication isn't "copy-paste your prompts." it's that causal attention creates a structural blind spot we've been ignoring. the first tokens in your prompt are always the least informed.
and for non-reasoning tasks, a second pass is the cheapest fix available.
tweet