home.social

#onyxaudit — Public Fediverse posts

Live and recent posts from across the Fediverse tagged #onyxaudit, aggregated by home.social.

fetched live
  1. "New mode makes GPT-5.6 Sol work at 14x the speed."
    It's a serving tier, not a model change. Up to 750 output tokens/sec on Cerebras hardware, limited preview.

    The 14x and the Claude comparison both come from Cerebras, whose chip is the product. GPT-5.6 Sol tested with Codex on 10 July, Claude Fable 5 with Claude Code on 13-15 July. Different harnesses, different days.

    Cerebras reported Q2 the day before.

    #AI #OnyxAudit

  2. "New mode makes GPT-5.6 Sol work at 14x the speed."
    It's a serving tier, not a model change. Up to 750 output tokens/sec on Cerebras hardware, limited preview.

    The 14x and the Claude comparison both come from Cerebras, whose chip is the product. GPT-5.6 Sol tested with Codex on 10 July, Claude Fable 5 with Claude Code on 13-15 July. Different harnesses, different days.

    Cerebras reported Q2 the day before.

    #AI #OnyxAudit

  3. "New mode makes GPT-5.6 Sol work at 14x the speed."
    It's a serving tier, not a model change. Up to 750 output tokens/sec on Cerebras hardware, limited preview.

    The 14x and the Claude comparison both come from Cerebras, whose chip is the product. GPT-5.6 Sol tested with Codex on 10 July, Claude Fable 5 with Claude Code on 13-15 July. Different harnesses, different days.

    Cerebras reported Q2 the day before.

    #AI #OnyxAudit

  4. "New mode makes GPT-5.6 Sol work at 14x the speed."
    It's a serving tier, not a model change. Up to 750 output tokens/sec on Cerebras hardware, limited preview.

    The 14x and the Claude comparison both come from Cerebras, whose chip is the product. GPT-5.6 Sol tested with Codex on 10 July, Claude Fable 5 with Claude Code on 13-15 July. Different harnesses, different days.

    Cerebras reported Q2 the day before.

    #AI #OnyxAudit

  5. "New mode makes GPT-5.6 Sol work at 14x the speed."
    It's a serving tier, not a model change. Up to 750 output tokens/sec on Cerebras hardware, limited preview.

    The 14x and the Claude comparison both come from Cerebras, whose chip is the product. GPT-5.6 Sol tested with Codex on 10 July, Claude Fable 5 with Claude Code on 13-15 July. Different harnesses, different days.

    Cerebras reported Q2 the day before.

    #AI #OnyxAudit

  6. X open-sourced the code that limits reach on the For You timeline, and shipped a tool showing you your visibility labels.
    I read the reply-spam flow, the exemption for government (grey badge) and high-reputation accounts skips the label, but the ranking score of 0.0 is still written afterwards, unconditionally.
    The tool shows labels.

    So the exempt accounts are the ones whose demotion the transparency tool cannot display.

    Full piece, evidence tiered: below reply
    #X #DSA #transparency #OnyxAudit

  7. X open-sourced the code that limits reach on the For You timeline, and shipped a tool showing you your visibility labels.
    I read the reply-spam flow, the exemption for government (grey badge) and high-reputation accounts skips the label, but the ranking score of 0.0 is still written afterwards, unconditionally.
    The tool shows labels.

    So the exempt accounts are the ones whose demotion the transparency tool cannot display.

    Full piece, evidence tiered: below reply
    #X #DSA #transparency #OnyxAudit

  8. X open-sourced the code that limits reach on the For You timeline, and shipped a tool showing you your visibility labels.
    I read the reply-spam flow, the exemption for government (grey badge) and high-reputation accounts skips the label, but the ranking score of 0.0 is still written afterwards, unconditionally.
    The tool shows labels.

    So the exempt accounts are the ones whose demotion the transparency tool cannot display.

    Full piece, evidence tiered: below reply
    #X #DSA #transparency #OnyxAudit

  9. X open-sourced the code that limits reach on the For You timeline, and shipped a tool showing you your visibility labels.
    I read the reply-spam flow, the exemption for government (grey badge) and high-reputation accounts skips the label, but the ranking score of 0.0 is still written afterwards, unconditionally.
    The tool shows labels.

    So the exempt accounts are the ones whose demotion the transparency tool cannot display.

    Full piece, evidence tiered: below reply
    #X #DSA #transparency #OnyxAudit

  10. X open-sourced the code that limits reach on the For You timeline, and shipped a tool showing you your visibility labels.
    I read the reply-spam flow, the exemption for government (grey badge) and high-reputation accounts skips the label, but the ranking score of 0.0 is still written afterwards, unconditionally.
    The tool shows labels.

    So the exempt accounts are the ones whose demotion the transparency tool cannot display.

    Full piece, evidence tiered: below reply
    #X #DSA #transparency #OnyxAudit

  11. @protonprivacy VPN's pricing page assigns a variant in the HTML, then draws your price from it. @WindScribe published the tag first. Pairing the letter to the price was not done.

    Eight readings, fixed conditions.

    A: €2.99, 70% off, billed €71.76
    B: €3.49, 65% off, billed €83.76

    Sequence: B, A, B, A, B across 74 minutes. Stale content changes once; it does not alternate.

    onyxdigital.bearblog.dev/proto

    #privacy #Proton #ABtesting #OnyxAudit

  12. Proton VPN's pricing page assigns a variant in the HTML, then draws your price from it. @WindScribe published the tag first. Pairing the letter to the price was not done.

    Eight readings, fixed conditions.

    A: €2.99, 70% off, billed €71.76
    B: €3.49, 65% off, billed €83.76

    Sequence: B, A, B, A, B across 74 minutes. Stale content changes once; it does not alternate.

    onyxdigital.bearblog.dev/proto

    #privacy #Proton #ABtesting #OnyxAudit

  13. Proton VPN's pricing page assigns a variant in the HTML, then draws your price from it. @WindScribe published the tag first. Pairing the letter to the price was not done.

    Eight readings, fixed conditions.

    A: €2.99, 70% off, billed €71.76
    B: €3.49, 65% off, billed €83.76

    Sequence: B, A, B, A, B across 74 minutes. Stale content changes once; it does not alternate.

    onyxdigital.bearblog.dev/proto

    #privacy #Proton #ABtesting #OnyxAudit

  14. Proton VPN's pricing page assigns a variant in the HTML, then draws your price from it. @WindScribe published the tag first. Pairing the letter to the price was not done.

    Eight readings, fixed conditions.

    A: €2.99, 70% off, billed €71.76
    B: €3.49, 65% off, billed €83.76

    Sequence: B, A, B, A, B across 74 minutes. Stale content changes once; it does not alternate.

    onyxdigital.bearblog.dev/proto

    #privacy #Proton #ABtesting #OnyxAudit

  15. Proton VPN's pricing page assigns a variant in the HTML, then draws your price from it. @WindScribe published the tag first. Pairing the letter to the price was not done.

    Eight readings, fixed conditions.

    A: €2.99, 70% off, billed €71.76
    B: €3.49, 65% off, billed €83.76

    Sequence: B, A, B, A, B across 74 minutes. Stale content changes once; it does not alternate.

    onyxdigital.bearblog.dev/proto

    #privacy #Proton #ABtesting #OnyxAudit

  16. Every attack model assumes someone is there. Actor, target, live connection.
    Take the person out and most of our tooling stops making sense.
    Exploits rot; that is their most reliable property. Semantic instructions do not. An instruction in language depends on nothing except something downstream being able to read it, and that capability is not going backwards.
    A thought experiment, labelled as one.
    onyxdigital.bearblog.dev/the-asynchronous-author/
    #AI #cybersecurity #promptinjection #OnyxAudit

  17. Every attack model assumes someone is there. Actor, target, live connection.
    Take the person out and most of our tooling stops making sense.
    Exploits rot; that is their most reliable property. Semantic instructions do not. An instruction in language depends on nothing except something downstream being able to read it, and that capability is not going backwards.
    A thought experiment, labelled as one.
    onyxdigital.bearblog.dev/the-asynchronous-author/
    #AI #cybersecurity #promptinjection #OnyxAudit

  18. Every attack model assumes someone is there. Actor, target, live connection.
    Take the person out and most of our tooling stops making sense.
    Exploits rot; that is their most reliable property. Semantic instructions do not. An instruction in language depends on nothing except something downstream being able to read it, and that capability is not going backwards.
    A thought experiment, labelled as one.
    onyxdigital.bearblog.dev/the-asynchronous-author/
    #AI #cybersecurity #promptinjection #OnyxAudit

  19. Every attack model assumes someone is there. Actor, target, live connection.
    Take the person out and most of our tooling stops making sense.
    Exploits rot; that is their most reliable property. Semantic instructions do not. An instruction in language depends on nothing except something downstream being able to read it, and that capability is not going backwards.
    A thought experiment, labelled as one.
    onyxdigital.bearblog.dev/the-asynchronous-author/
    #AI #cybersecurity #promptinjection #OnyxAudit

  20. Every attack model assumes someone is there. Actor, target, live connection.
    Take the person out and most of our tooling stops making sense.
    Exploits rot; that is their most reliable property. Semantic instructions do not. An instruction in language depends on nothing except something downstream being able to read it, and that capability is not going backwards.
    A thought experiment, labelled as one.
    onyxdigital.bearblog.dev/the-asynchronous-author/
    #AI #cybersecurity #promptinjection #OnyxAudit

  21. Six security outlets reported 868 npm packages compromised today. The source page now reads 434. Same timestamp, same 1,381 versions, exactly half the packages.

    I went to check the original. Nobody had archived it. Mine was the first snapshot in existence, six hours after the figure entered circulation.

    Five different counts for one event inside nine hours.

    onyxdigital.bearblog.dev/first
    #npm #supplychain #ShaiHulud #infosec #OnyxAudit