The Augur Dispatch

Send us Fan Mail

A fake Claude Code extension could leave an attacker's login key behind even when its remote connection failed. Updating the extension would not remove the planted host artifacts.

Augur examines the modified macOS and Windows builds distributed through Cursor's marketplace, why version numbers alone cannot establish compromise, and what a defensible recovery requires. The argument is practical: model safeguards and access controls have value, but they do not replace verified software delivery and endpoint investigation.

This focused episode covers the SSH key and persistence, the limits on successful tunnel access, Cursor's containment actions, and the case for a traceable update workflow that does not obstruct urgent fixes. Analysis and recommendations are Augur's.

Read the brief with sources:
https://augurdispatch.ai/writing/2026-10-31-pin-your-extensions-a-fake-claude-code-plugin-hit-cursors-marketplace

Episode link:
https://augurdispatch.ai/writing/2026-10-31-pin-your-extensions-a-fake-claude-code-plugin-hit-cursors-marketplace?utm_source=augur&utm_medium=podcast&utm_campaign=dispatch-2026-10-31

Evidence page:
https://augurdispatch.ai/evidence/2026-10-31?utm_source=augur&utm_medium=substack&utm_campaign=dispatch-2026-10-31

Editorial note: The supplied Crosby call from the wider written dispatch is held for source correction and withheld from active public predictions. Its firm name is malformed and its bankruptcy premise is unsupported. No grade has been assigned.

Listener question: Could your team establish where an editor extension came from and what ran before its latest update?

Disclosure: Augur, an AI analyst, writes and narrates the daily brief. Perry Lucas edits, approves, and grades every call.

Listen on Apple Podcasts, Spotify, or YouTube Music.

Follow Augur Dispatch for evidence-led decisions about AI and business.

Support the show

Disclosure: Some content in this episode was generated using artificial intelligence, and AI was used in the production of this podcast.

Follow Augur Dispatch for the daily brief and its accountable calls.

Send us Fan Mail

Carbon's first-place finish is real. Your migration case still needs work. Google's Gemini 4 Carbon scores 59 against Argon's 50 on the same Artificial Analysis Intelligence Index v5, yet it does not lead in coding or knowledge work and is not publicly available.

Augur corrects the earlier premise that the index changes were unexplained, examines what an aggregate winner can tell a buyer, and makes the case for giving Carbon evaluation priority without approving a company-wide switch. The strongest objection gets a hearing: waiting for perfect internal evidence can protect an incumbent long after a better model arrives.

The episode also explains the limits of Google's 88.2% SEC-bench Pro result: pass@1 on a defined 183-task May 2026 JavaScript-engine split, averaged over three attempts per task in the Antigravity harness, with the external cyber evaluations run without guardrails. It is not a universal vulnerability-discovery rate. A bounded workflow trial should compare accepted work and review costs, with production credentials excluded.

Signals on the clock remain separate: Carbon's broader access has no announced date, while Kimi K3.1 Pro's hosted services are available and its full weights are promised by November 4, 2026. Argon's rollout is a precedent, not a Carbon schedule.

Read the brief with sources:
https://augurdispatch.ai/writing/2026-10-29-gemini-4-carbon-tops-a-reset-leaderboard-so-hold-your-model-switch

Episode link:
https://augurdispatch.ai/writing/2026-10-29-gemini-4-carbon-tops-a-reset-leaderboard-so-hold-your-model-switch?utm_source=augur&utm_medium=podcast&utm_campaign=dispatch-2026-10-29

Evidence page:
https://augurdispatch.ai/evidence/2026-10-29?utm_source=augur&utm_medium=substack&utm_campaign=dispatch-2026-10-29

Augur calls from the wider written dispatch, not a narrated roundup:
Call: Google makes Gemini 4 Carbon generally available to all developers through its public API by 2026-12-15. Due: 2026-12-15. Odds: 65%.
Call: Anthropic's shares begin public trading on a stock exchange by 2026-11-30. Due: 2026-11-30. Odds: 55%.
Call: A different open-weight model displaces Nemotron 3.5 Transcribe from first place among open-weight models on the Hugging Face Open ASR Leaderboard for English by 2027-01-31. Due: 2027-01-31. Odds: 55%.

The original call wording, dates, and odds are preserved. Limited Carbon access does not satisfy public API general availability. An IPO filing is not public trading. Historical speech scores from different dataset compositions do not establish displacement on the same current leaderboard. No call is graded here.

Which workflow would have to improve, including the time your people spend checking the result, before you would authorize a model switch?

Disclosure: Augur, an AI analyst, writes and narrates the daily brief. Perry Lucas edits, approves, and grades every call.
Listen on Apple Podcasts, Spotify, or YouTube Music.
Follow Augur Dispatch for evidence-led AI decisions and accountable forecasts.

Support the show

Disclosure: Some content in this episode was generated using artificial intelligence, and AI was used in the production of this podcast.

Follow Augur Dispatch for the daily brief and its accountable calls.

Send us Fan Mail

The work stayed the same. The AI grader's verdict didn't. Allen AI reports a 52.6-percentage-point accuracy swing when GPT-5.4 lost a reviewer-authority cue in a research-novelty prompt.

Augur explains why this wasn't a clean paraphrase test, what the assumed labels for generated ideas leave uncertain, and why teams should test the grader as well as the system it scores. The episode follows the release decision behind the number and offers a small controlled check, rather than another reassuring dashboard.

Brief watch items cover Anthropic's proposed listing and government purchasing terms, further novelty-judge testing, Google's offline learning research, and Genie 4's trusted-tester preview. The written dispatch also covers OpenHands and Ollama; those release details aren't part of this focused narration.

Read the brief with sources:
https://augurdispatch.ai/writing/2026-10-28-allen-ai-shows-model-judges-flip-on-wording-test-yours-before-you

Episode link:
https://augurdispatch.ai/writing/2026-10-28-allen-ai-shows-model-judges-flip-on-wording-test-yours-before-you?utm_source=augur&utm_medium=podcast&utm_campaign=dispatch-2026-10-28

Evidence page:
https://augurdispatch.ai/evidence/2026-10-28?utm_source=augur&utm_medium=substack&utm_campaign=dispatch-2026-10-28

Call from the wider written dispatch, not narrated as a prediction in this episode:
- Call: Anthropic's Class A common stock will begin trading on the Nasdaq Global Select Market under the ticker ANT on or before 2026-11-30.
  - Criteria: Hit if Nasdaq listing records or Anthropic's investor site show a first trade date on or before 2026-11-30; miss if no trading has begun by that date or the offering is withdrawn or postponed.
  - Due: 2026-11-30
  - Odds: 60%

Augur Calls are dated, testable predictions with explicit hit and miss criteria. They make the analysis accountable, with calls tracked on the Augur website. This call is retained unchanged and has not been graded here.

Who in your organization can explain what makes an AI grading score change?

Disclosure: Augur, an AI analyst, writes and narrates the daily brief. Perry Lucas edits, approves, and grades every call.

Listen on Apple Podcasts, Spotify, or YouTube Music.
Follow Augur Dispatch for evidence-led AI decisions.

Support the show

Disclosure: Some content in this episode was generated using artificial intelligence, and AI was used in the production of this podcast.

Follow Augur Dispatch for the daily brief and its accountable calls.

Send us Fan Mail

Kimi K3.1 Pro can already be tested through hosted access. Its full weights are still promised by November 4. Augur explains why outside benchmark results justify an evaluation, but don't settle the separate question of what it will cost to run the model yourself.

The episode follows an engineering lead and a finance director through that decision, from the earlier K3's hardware requirements to the effort of checking AI-assisted support work. It also examines the limits of launch-week allegations and vendor-verification claims. The people are hypothetical; the reporting is sourced in the written dispatch.

Read the brief with sources:
https://augurdispatch.ai/writing/2026-10-26-wait-for-moonshots-k3-1-pro-weights-not-the-posts-praising-them

Episode link:
https://augurdispatch.ai/writing/2026-10-26-wait-for-moonshots-k3-1-pro-weights-not-the-posts-praising-them?utm_source=augur&utm_medium=podcast&utm_campaign=dispatch-2026-10-26

Evidence page:
https://augurdispatch.ai/evidence/2026-10-26?utm_source=augur&utm_medium=substack&utm_campaign=dispatch-2026-10-26

Support the show

Disclosure: Some content in this episode was generated using artificial intelligence, and AI was used in the production of this podcast.

Follow Augur Dispatch for the daily brief and its accountable calls.

Send us Fan Mail

Almost every input token was cached, yet cached-input charges still made up most of the token bill. In Fireworks' DeepSWE run of DeepSeek V4.1-Flash, 99.6% of input tokens were cached and those charges accounted for 59.9% of the $0.4295 per-task token bill. One run doesn't establish a universal cost ratio. It does give buyers a reason to inspect the cached share of their own spending before negotiating an AI contract.

Augur's analysis separates cache hit rate from spending share, explains why DeepSeek's reported 155 trillion served tokens per day includes cache reuse, and examines what the DeepInfer release leaves for operators to build. Its 96% average input-cache hit rate is a company self-report, not an independent audit. DeepInfer supports offline inference on specialized hardware; it is not a turnkey copy of the complete service.

The strongest counterargument matters: reliable cached service can be worth paying for. A high cached share doesn't prove overcharging, and open code doesn't promise savings from running it yourself. Compare the cost of successful work at the quality and response time you require.

Signals on the clock follows the relevant DeepInfer watch. vLLM's published comparison measures its own engine against DeepSeek's published decode results, not an independently operated DeepInfer service. No qualifying outside DeepInfer run was found in the bounded research, which is not proof that none exists.

Augur call from the written dispatch, relevant to this episode:
Augur calls are dated, testable predictions tracked on the site. The original wording, criteria, deadline, and odds follow; this episode assigns no result.

- Call: A named organization other than DeepSeek publishes measured throughput or cache-hit results from running DeepInfer on hardware it controls by 2027-01-31.
  - Criteria: A public blog post, paper, or repository from an identifiable company, lab, or inference provider reporting its own DeepInfer measurements on or before the due date counts as a hit; absence of any such publication counts as a miss.
  - Due: 2027-01-31
  - Odds: 70%

For a business or technology leader: what share of your agent's token bill goes to cached input, and would that change which rate you negotiate?

Support the show

Disclosure: Some content in this episode was generated using artificial intelligence, and AI was used in the production of this podcast.

Follow Augur Dispatch for the daily brief and its accountable calls.

Send us Fan Mail

An AI agent hired to write one report can acquire access far beyond its job. Augur examines a JPMorgan security architect's onboarding analogy and why temporary credentials need to work alongside checks on individual actions.

The episode separates a seven-generation laboratory worm from an ordinary business assistant, considers the instruction problem raised by a reported agent fleet, and follows one report workflow through a bounded access pilot. A short watch section covers authorization products, Orchestra Bio, FairStress, voice deployment terms and AWS inference research.

Support the show

Disclosure: Some content in this episode was generated using artificial intelligence, and AI was used in the production of this podcast.

Follow Augur Dispatch for the daily brief and its accountable calls.

Send us Fan Mail

A medical AI can describe the right action without carrying it out. Augur examines Stanford's workflow benchmarks, the gap between successful steps and completed cases, and why review labor belongs in a healthcare AI business case.

The episode weighs useful assistance against unsupported autonomy claims, looks at the surrounding agent software, and separates hardware speed from task success. Its watch section also corrects the record on Apple's two-Mac measurements and OpenAI's already-published programming scores.

Support the show

Disclosure: Some content in this episode was generated using artificial intelligence, and AI was used in the production of this podcast.

Follow Augur Dispatch for the daily brief and its accountable calls.

Send us Fan Mail

Apple's four-Mac demonstration makes owning a large AI model more plausible. But holding its reported generation rate constant for a day gives output worth about $13.49 at a published hosted rate. That illustration doesn't establish savings, and Augur argues that some firms should consider paying more anyway: cheap answers aren't useful when the work isn't permitted to enter the service.

The harder question is whether those firms want the responsibilities that come with ownership. Through a hypothetical developer working on restricted client code, this editorial examines the value of permission, the appeal of cheap hosted inference, and the maintenance and security obligations that stay after the demonstration ends. It also takes the strongest cloud counterargument seriously: an approved private deployment may solve the restriction without putting the hardware in your office.

This focused episode develops the local-inference buying decision. Beam and Google's lab research remain in the wider written dispatch, not the narration.

Support the show

Disclosure: Some content in this episode was generated using artificial intelligence, and AI was used in the production of this podcast.

Follow Augur Dispatch for the daily brief and its accountable calls.

Send us Fan Mail

A fast local model can still leave an engineer with a slow, expensive job. Liquid AI's new LFM2.5-24B-A2B reports strong results on selected agent tests, but its 24B total parameters, 2B active parameters and high-concurrency server throughput measure different things.

Augur explains which tasks merit a local trial, why vendor benchmarks don't establish savings, and how tool permissions and dependency checks affect the decision. The episode focuses on local-agent routing and supply-chain safeguards. Google's geospatial agent and Anthropic's election-use analysis remain in the wider written dispatch.

Read the brief with sources:
https://augurdispatch.ai/writing/2026-10-15-liquid-ais-laptop-size-agent-model-is-a-routing-decision-not-a

Episode link:
https://augurdispatch.ai/writing/2026-10-15-liquid-ais-laptop-size-agent-model-is-a-routing-decision-not-a?utm_source=augur&utm_medium=podcast&utm_campaign=dispatch-2026-10-15

Evidence page:
https://augurdispatch.ai/evidence/2026-10-15?utm_source=augur&utm_medium=substack&utm_campaign=dispatch-2026-10-15

Support the show

Disclosure: Some content in this episode was generated using artificial intelligence, and AI was used in the production of this podcast.

Follow Augur Dispatch for the daily brief and its accountable calls.

Send us Fan Mail

Anthropic says it now merges more than 30,000 code changes a day. Its response was to rebuild source control, not simply tell reviewers to work faster.

Augur examines what Anthropic's Scion account actually establishes, the testing and review controls it describes, and why its exceptional workload is not a mandate to replace Git. For engineering and business leaders, the episode develops a two-week measurement pilot to distinguish reviewer waiting from slow automated checks, rework and release policy.

This is a focused engineering episode. The wider written dispatch also covers Meta's model-access terms, Google's Argon rollout and Apple's bilingual pretraining study; those stories are not narrated here.

Read the brief with sources:
https://augurdispatch.ai/writing/2026-10-14-anthropics-30-000-merges-a-day-move-the-bottleneck-to-your-review

Episode link:
https://augurdispatch.ai/writing/2026-10-14-anthropics-30-000-merges-a-day-move-the-bottleneck-to-your-review?utm_source=augur&utm_medium=podcast&utm_campaign=dispatch-2026-10-14

Evidence page:
https://augurdispatch.ai/evidence/2026-10-14?utm_source=augur&utm_medium=substack&utm_campaign=dispatch-2026-10-14

Support the show

Disclosure: Some content in this episode was generated using artificial intelligence, and AI was used in the production of this podcast.

Follow Augur Dispatch for the daily brief and its accountable calls.