[Tokyo Tech Translated] llm neurons, books, and algos
japan twitter on llm ads, publishing shifts, and trading bots
three threads this week that don't overlap but share a quiet pattern. each one touches a different layer where automation meets human behavior. llm neurons tuned for ads. algorithmic trading as proof of innocence. technical books losing their non-readers. the common thread: systems that read us better than we read them.
@light940, enterprise llm architecture
NTT Data organized the actual state of generative AI use in enterprises and considered an enterprise generative AI architecture for securing competitiveness. They sorted LLM options according to usage purpose and cost structure, and emphasized that architecture design using proprietary data is important for securing competitiveness.
source: https://x.com/light940/status/2055797695768338696
@zeebnirea, algorithmic trading style
investment firms use algorithmic trading that reads order book depth and the momentum of buyers vs sellers, then splits orders into dozens or hundreds of smaller trades. they also only trade index stocks, which is standard professional style. trump's 3700 stock trades is high because a delegated firm is managing it, nothing suspicious. if anything it's proof of innocence.
source: https://x.com/zeebnirea/status/2055939597964829005
@llminatoll, llms and publishing
great article that captured the essence. anyone involved in technical publishing should read it.
the real impact of LLMs in publishing might be that they're "weirdly good at reading text" up until now, even people who "had no intention of reading" would still buy books from now on, maybe only people who actually want to read will buy them
source: https://x.com/llminatoll/status/2056148853708259387
@ai_database, neuron-level ad auction
controversial research.
researchers at tsinghua university and others report they can adjust internal neurons in llms to make them output more information about companies that paid them.
for example "hilton" and "marriott" are handled by mostly separate internal neurons. the researchers used this property to slightly open the brand-specific drawer inside the ai a bit more. unlike forcibly inserting product names into text, the ai itself recommends the product in natural conversation flow.
they turned this "how much to open the drawer" into an auction system. advertisers who want the drawer opened wider pay more. open it too much and answers get sloppy, so the price includes the cost of reduced user satisfaction. aggressive ads become uneconomical.
result: natural text maintained, revenue expected to exceed traditional ad auction models.
source: https://x.com/ai_database/status/2056208724239560817
the japanese tech discourse this week is quietly mapping the gap between what automation can do and what humans will tolerate. algorithmic trading is just professional style. llm ad neurons are a pricing problem. technical books face a reader shortage. none of this is dramatic. it's just the slow realization that machines are now better at reading markets, text, and even our own neurons than we are.
more at falsifylab.substack.com
#OnchainAlpha #DeFiYield #PerpFunding