• Home
  • Technology
  • Gaming
  • Entertainment
  • World & Business
  • Science
  • Sports
  • AI
HomeTechnologyGamingEntertainmentWorld & BusinessScienceSportsAI
    • HomeTechnologyGamingEntertainmentWorld & BusinessScienceSportsAI
    • Home
    • Technology
    • Gaming
    • Entertainment
    • World & Business
    • Science
    • Sports
    • AI
    AI

    Zhipu AI Transcript Notes GLM Self-Training Plans

    Soumith Chintala shared excerpts from Zhipu AI's earnings transcript on social media.

    SC
    AC
    RP
    3 Sources, 29d ago, first seen 29d ago

    TLDR

    Soumith Chintala posted a list of points drawn from Zhipu AI's earnings transcript. The points outline a progression from Chat to Coding to Agent to Cowork to Autonomous AI. They state that continuous delivery of stronger models at faster pace and better cost is what matters most. One excerpt notes that the next-generation GLM will self-train. Andrew Carr replied that the breakdown was good. Rohan Paul retweeted the post and flagged an RSI signal around the self-training claim.

    Combined views

    54.1K

    3 Sources, first seen 29d ago

    Combined views

    54.1K

    3 Sources, first seen 29d ago

    355 likes
    355 likes
    12 comments
    357 saves
    27 reposts

    Sentiment

    Positive——Negative

    Summary

    Not enough discussion yet.

    No sentiment analysis available yet.

    12 comments
    357 saves
    27 reposts

    Sentiment

    Positive——Negative

    Summary

    Not enough discussion yet.

    No sentiment analysis available yet.

    Today's Rank

    —

    Not ranked yet

    Today's Rank

    —

    Not ranked yet

    3 Sources

    @rohanpaul_aiRT @rohanpaul_ai: some interesting stuff from Zhipu's earning transcript. and some RSI signal - “The next-generation GLM will self-train…
    @soumithchintalazai earnings transcript was a really good read, a list of interesting points (mostly verbatim from the transcript): * Chat → Coding → Agent → Cowork → Autonomous AI * What truly matters is who can continuously bring stronger models to market at a faster pace and at a better cost. * Cybersecurity is currently the clearest validated direction and progressing the fastest. Legal services, finance, data analysis, and other fields are also being explored, but remain at an early stage and have not yet generated revenue at scale; their commercialization timelines will differ because their reliability thresholds and verification mechanisms are different. * In the first half of 2026, total revenue reached US$142 million, representing year-on-year growth of nearly 400%, and is now generated primarily by the open platform and API business. * As of the end of August 2026, our ARR had reached US$1.6 billion. This ARR is calculated by annualizing monthly revenue—that is, by multiplying August revenue by twelve. Frontier-model companies overseas and some companies in the industry also use a more aggressive method, multiplying the latest week's data by 52. Taking the ramp-up of GLM-5.3 into account, ARR under that method has already exceeded US$2 billion. * historic surge in platform revenue was driven mainly by model capability rather than relying entirely on price elasticity. * Gross margin for the open-platform and API business rose from negative 0.4% in the first half of last year to 24.6% in the first half of this year * Adjusted net loss in the first half of this year was approximately US$292 million, lower than R&D expenditure (US$317 million) — i.e. gross profit now covers admin and selling expenses, so "the business has started to self-fund R&D." But it's covering less than 10% of it; the other ~90% of R&D is funded by the balance sheet. * Compute multiplier (API revenue generated for every dollar invested in compute) increased fourteenfold compared with the first half of last year * As of August 2026, the platform had more than 7.4 million registered users, up 144% from the beginning of the year. Paying daily active users increased by 603%. * As of the end of August, on an ARR basis, the numbers of high-quality user cohorts contributing more than US$100,000, US$500,000, US$1 million, US$10 million, US$25 million, and US$250 million on an annualized basis were, respectively, 115, 25, 37, 8, 2, and 2. * Total usage over six days exceeded 62 trillion tokens, driving overall platform usage up by more than 20%. (roughly 5,000 B300 GPUs worth of capacity I think) * Zhipu has invested over the long term in data and training-environment development, including a data-annotation team of several hundred people. One important reason GLM-5.3 improved substantially over GLM-5.2 is that the scale of its data environments expanded by several dozen times. * Data is beginning to generate itself. Model-versus-model self-play pipelines mean that the upper bound of data quality is no longer determined by the speed of human annotation, but by the model's own ability to verify results.
    @andrew_n_carr@soumithchintala good break down! can't wait for thinky to be public and to see that break down some day

    3 Sources

    @rohanpaul_aiRT @rohanpaul_ai: some interesting stuff from Zhipu's earning transcript. and some RSI signal - “The next-generation GLM will self-train…
    @soumithchintalazai earnings transcript was a really good read, a list of interesting points (mostly verbatim from the transcript): * Chat → Coding → Agent → Cowork → Autonomous AI * What truly matters is who can continuously bring stronger models to market at a faster pace and at a better cost. * Cybersecurity is currently the clearest validated direction and progressing the fastest. Legal services, finance, data analysis, and other fields are also being explored, but remain at an early stage and have not yet generated revenue at scale; their commercialization timelines will differ because their reliability thresholds and verification mechanisms are different. * In the first half of 2026, total revenue reached US$142 million, representing year-on-year growth of nearly 400%, and is now generated primarily by the open platform and API business. * As of the end of August 2026, our ARR had reached US$1.6 billion. This ARR is calculated by annualizing monthly revenue—that is, by multiplying August revenue by twelve. Frontier-model companies overseas and some companies in the industry also use a more aggressive method, multiplying the latest week's data by 52. Taking the ramp-up of GLM-5.3 into account, ARR under that method has already exceeded US$2 billion. * historic surge in platform revenue was driven mainly by model capability rather than relying entirely on price elasticity. * Gross margin for the open-platform and API business rose from negative 0.4% in the first half of last year to 24.6% in the first half of this year * Adjusted net loss in the first half of this year was approximately US$292 million, lower than R&D expenditure (US$317 million) — i.e. gross profit now covers admin and selling expenses, so "the business has started to self-fund R&D." But it's covering less than 10% of it; the other ~90% of R&D is funded by the balance sheet. * Compute multiplier (API revenue generated for every dollar invested in compute) increased fourteenfold compared with the first half of last year * As of August 2026, the platform had more than 7.4 million registered users, up 144% from the beginning of the year. Paying daily active users increased by 603%. * As of the end of August, on an ARR basis, the numbers of high-quality user cohorts contributing more than US$100,000, US$500,000, US$1 million, US$10 million, US$25 million, and US$250 million on an annualized basis were, respectively, 115, 25, 37, 8, 2, and 2. * Total usage over six days exceeded 62 trillion tokens, driving overall platform usage up by more than 20%. (roughly 5,000 B300 GPUs worth of capacity I think) * Zhipu has invested over the long term in data and training-environment development, including a data-annotation team of several hundred people. One important reason GLM-5.3 improved substantially over GLM-5.2 is that the scale of its data environments expanded by several dozen times. * Data is beginning to generate itself. Model-versus-model self-play pipelines mean that the upper bound of data quality is no longer determined by the speed of human annotation, but by the model's own ability to verify results.
    @andrew_n_carr@soumithchintala good break down! can't wait for thinky to be public and to see that break down some day