The AI Enthusiast.

See what your agents did. Claude's agents now report every step to the monitoring tools your ops team already runs. #

OpenTelemetry public beta October 14, 2026 · Uber CISO essay October 13 · Nadella interview October 15

On October 14, Anthropic put OpenTelemetry (OTel) traces from Claude Managed Agents, the Messages API and the Claude SDKs into public beta. Every agent turn, model call, tool use and MCP call can now arrive in any tool that accepts the OTLP standard, with latency, token counts and IDs, laid out with the OpenTelemetry GenAI semantic conventions, so Claude spans line up with spans from other providers.

Trusting an agent with more access starts with tracing its model calls, tool use, and MCP server calls. Review its traces, then widen its access as it proves reliable.Anthropic, October 14, 2026

Why this week

Three things landed within 72 hours. On October 13, Uber's CISO Philip Martin argued that AI agents need "the equivalent" of the insider-threat program every company already runs for employees. On October 15, NBC published an interview in which Satya Nadella called all AI agents an "insider risk". And Anthropic shipped the plumbing that makes watching agents routine.

What a trace looks like

Your first trace in about 15 lines

From the docs. Install with pip install --upgrade anthropic opentelemetry-sdk; the console prints one anthropic.messages.create span. Then swap the console exporter for an OTLP exporter pointed at your platform.

from anthropic import Anthropic
from opentelemetry import trace
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import ConsoleSpanExporter, SimpleSpanProcessor

# Register a tracer provider that prints each span to the console.
provider = TracerProvider()
provider.add_span_processor(SimpleSpanProcessor(ConsoleSpanExporter()))
trace.set_tracer_provider(provider)

client = Anthropic()
message = client.messages.create(
    model="claude-opus-5-5",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello, Claude"}],
)
print(message.content)

Set ANTHROPIC_OPEN_TELEMETRY_TRACES_CONTENT_MODE=content only if you want prompts and responses in your spans. Pin the SDK version: the docs warn that span names and options can change during the beta. Each span holds up to 0.75 MB of content by default (about a 200k-token context); a full 1M-token context needs about 4 MB.

Where traces can go

Launch partners: Grafana Cloud, Honeycomb, Sentry, Splunk (traces beside logs and metrics); Arize AI, Braintrust, Datadog, Langfuse, LangSmith, New Relic, PostHog (turn traces into evals); and Databricks Unity Gateway (token usage by team). Ramp says its "Reflect" system scores Langfuse traces and has cut token usage 10–20% and made some agent sessions 30% faster (customer figures from Anthropic's post, self-reported).

Claude Code joins in. This week's releases added GenAI semantic-convention attributes to its spans (gen_ai.conversation.id, gen_ai.tool.name, error.type; 2.1.300) and classifier_verdict, permission_mode and classifier_category on tool_decision events, so you can tell an auto-mode classifier block from a permission rule (2.1.301).

The insider-threat model, from Uber

Martin's architecture has three layers: an LLM gateway that sets a perimeter on AI use; Uber's newly open-sourced AI Detection and Response (ADR), which gathers real-time telemetry and can act inside agent harnesses; and Anthropic's Enterprise Frontier Safeguards classifier output as one more signal. "Every company already runs an insider threat program for its people. Nobody argues about that."

We tried the chatbot that answers in a tenth of a second. ChatJimmy runs on a chip with the model etched into it, and AMD is buying the company that made it. #

Our hands-on test October 15, 2026 · AMD agreement announced August 6, 2026 (before this window) · Taalas launch post February 19, 2026

ChatJimmy (chatjimmy.ai) is the public demo of Taalas's HC1, a chip with Llama 3.1 8B hard-wired into the silicon. Taalas claims 17,000 tokens per second per user, "nearly 10X faster than the current state of the art," at 20x lower build cost and 10x less power (all self-reported). We tested it on Thursday night: every answer, including a full email, was on screen about 0.1 seconds after we pressed Enter. In August, AMD agreed to acquire Taalas; the price wasn't disclosed and, as of October 15, we found no report that the deal has closed.

How it works: the model is the computer

A normal AI server keeps the model's weights in memory (HBM) next to a GPU and streams them through the chip for every token. That trip, not the math, is the bottleneck, and it's why AI servers need HBM stacks, advanced packaging and liquid cooling. Taalas's answer is to build one chip per model, merging storage and compute "on a single chip, at DRAM-level density," with "no HBM, advanced packaging, 3D stacking, liquid cooling, high speed IO." The cost: you can't load a different model. Taalas says it can turn "a previously unseen model" into hardware "in only two months," and supports LoRA adapters and configurable context on top.

Hands-on: three questions, timed

We used ChatJimmy from a US home connection on October 15, around 11 PM ET, with headless Chrome, no login and no personal data. The deployment has a 6,144-token context. Answers appeared in one piece rather than streaming.

Most of the tenth of a second is network round trip: server-side generation was 6 to 23 milliseconds. The on-screen rate, about 14,300 tok/s, is below the 17K headline, likely because of short outputs and prompt processing. The answers are what you'd expect from a small, heavily quantized model, which Taalas itself warns "introduces some quality degradations."

The AMD deal, and why it matters now

On August 6, 2026, AMD said it had "reached a definitive agreement to acquire Taalas," subject to "customary closing conditions and regulatory approvals." No price, no closing date. AMD plans to "integrate Taalas' technology to deliver system-level solutions together with AMD Instinct™ GPUs." Its release never mentions hard-wiring, HC1 or ChatJimmy. Taalas, founded in 2023 in Toronto by former Tenstorrent founder Ljubisa Bajic, says it built HC1 with 24 people and "just $30M spent, of more than $200M raised." Its next chip, HC2, is meant to carry "a frontier LLM," with deployment "planned for winter."

We founded Taalas to rethink AI inference from the ground up by building the hardware around the model.Ljubisa Bajic, CEO, Taalas · AMD release, August 6, 2026

Jev 1.14 makes AI decisions in a tenth of a second. A "System One" model that answers with types and probabilities instead of text, and how we used it. #

Jev 1.14.0 ("JevDay") October 14, 2026 · $870M Series A October 9 · launch September 15

Jev, from San Francisco's TypeSafe AI, doesn't write text. You send it state (text or JSON) and typed questions: yes/no, pick-one, score-on-a-scale and, new this week, a bounding box. It returns typed answers with calibrated probabilities in about 70–500 ms, for $0.042 per million input tokens, with output free. On October 14 TypeSafe shipped Jev 1.14.0 as jev-preview, with better multilingual support and opt-in "Danger Zone" features. Five days earlier it raised $870M at a $7.5B valuation, led by Andreessen Horowitz. (Issue 24 covered Jev's September launch; this is the first major update, and the first time we've built with it.)

How it differs from a chat model

The name comes from Daniel Kahneman's fast "System 1" thinking. A chat model writes an answer one token at a time, and your code then parses it, validates it and retries when it's malformed. Jev evaluates every question in parallel against the same state and can only answer within the type you defined, so there is nothing to parse. TypeSafe trains it with what it calls RLCD, Reinforcement Learning for Calibrated Decisions: an answer given 0.8 should be right about 80% of the time. That's what makes a threshold in your code meaningful.

Our real use case: triaging failed API tests

In our Postman Local build (story 01), a collection run against SOAP services fails in different ways, and each needs a different response: an expired token means stop and log in again, a 503 means retry, a server exception means flag it and maybe page someone. Regexes over status codes and fault strings get most of the way and then rot. So the runner sends Jev one typed question set per failure, and acts only when the answer clears a confidence threshold. TypeSafe calls this a "smart if-statement."

This is the actual request our code builds for a SOAP 1.2 server fault from the fictional Acme Retail mock (abridged; the full criteria text is in the project's docs):

POST https://api.typesafe.ai/v1/systemone
Authorization: Bearer $TYPESAFE_API_KEY

{
  "model": "jev-preview",
  "state": {
    "request":  { "name": "GetOrder", "method": "POST" },
    "response": { "status_code": 500, "time_ms": 212 },
    "soap_fault": { "version": "1.2", "code": "soap:Receiver",
      "reason": "Internal error: NullReferenceException in OrderRepository::load()" }
  },
  "questions": {
    "category": { "type": "choice",
      "instructions": "Classify the most likely root cause of this failure…",
      "criteria": { "auth_expired": "…", "bad_input": "…", "env_misconfigured": "…",
                    "server_bug": "…", "transient_retry": "…" } },
    "should_retry": { "type": "noul", "instructions": "Retrying the same request is likely to succeed." },
    "page_oncall":  { "type": "noul", "instructions": "This warrants paging on-call now." }
  }
}

And the documented shape of what comes back: a choice with a probability for every option and a confidence, plus a noul P(yes) for each yes/no question.

{ "model": "jev-1.14.0",
  "answers": {
    "category":     { "type": "choice", "choice": "…", "probabilities": { "server_bug": p, … }, "confidence": c },
    "should_retry": { "type": "noul", "noul": p },
    "page_oncall":  { "type": "noul", "noul": p } },
  "usage": { "input_tokens": n, "output_tokens": n } }

What Jev actually said

Four live calls on October 16: the three failures in the Acme Retail collection run, plus a wrong password Irv sent by hand. Latency is measured from Postman Local, round trip, from Ohio.

The three scripted failures were easy, and Jev treated them that way. The wrong password is the interesting one. "Invalid login or password" really is ambiguous between an expired session and bad input, and Jev said so, splitting 79/21. One number, the threshold, turned that doubt into the right action: ask a person. Each call used about 860 to 1,130 input tokens, which at $0.042 per million is well under a hundredth of a cent.

Why typed and calibrated matters here: the category can only be one of five strings your code already handles, so there's no "JSON-ish" answer to repair, and the 0.8 threshold sets your automation rate against your error rate with one number you can tune from logged outcomes.

Speed and cost, by TypeSafe's numbers

What JevDay added

TypeSafe's own framing: "there is still way too much jaggedness for us to vouch for these for production use cases," and "expect most releases to be beautifully boring (like we think AI should be)." Competition arrived fast: OpenAI opened a Decisions API on GPT-6 Luna to everyone on October 6, and Microsoft released Decision-1 around October 9–12.

Small AI models may soon copy themselves onto smart TVs and ovens. Anthropic's red team measured how close they are, and what stops them. #

October 12, 2026 · Anthropic Frontier Red Team (William Zhou, Daniel Freeman)

Anthropic's Frontier Red Team says small open models, 30B parameters or less, "may be months away" from being able to self-replicate on commercial IoT devices such as smart ovens and off-brand Android TV boxes, with "a doubling time of days." The point: a worm no longer needs a data center. "The weights would become part of the malware payload."

The test bed was deliberately easy: 8 Google Cloud nodes running an Android 15 emulator with an unauthenticated ADB port. Setup is slow on weak hardware, 71–150 hours for Gemma4-26B-A4B unoptimized and 28–51 hours for Qwen3.5-9B, but it doesn't need root, because inference engines run in user space. Anthropic sizes the threat honestly: a Kimwolf-sized botnet of 2M+ devices would yield ~133,000 Gemma replicas and ~80,000 tokens/s, "at most about 40 H100s-worth of inference." The danger is persistence and spread, not raw compute.

Like coding agents in 2024, we're seeing signs of life for AI worms.Anthropic Frontier Red Team, October 12, 2026

Watch: how the networks covered it

3:58

AI researcher unpacks Anthropic's self-replication warningCBS News · Jonas Guan, U of Toronto

3:21

Anthropic warns AI models could soon copy themselves in small devicesABC News

7:26

What to know about new warning from AnthropicABC News · Oct 15

9:58

How self-replicating A.I. shifts the global A.I. raceFOX 35 Orlando · Oct 15

Anthropic says it cut model copying by 90 to 99%. How it stopped labs from distilling Claude, and what that costs developers. #

October 12, 2026 · Anthropic Policy · plus "unintended model actions," October 9

Distillation means training your own model on another model's outputs, especially its step-by-step reasoning. Anthropic says new defenses cut what distillers could harvest "by around 90 to 99%": once its classifier was on, reasoning tokens exposed to distillers were "consistently 10 to 100 times lower," and on some days too small to measure. It has banned "tens of thousands of accounts," often "in less than two hours" after a campaign starts.

Who was copying

Anthropic names names. Zhipu distilled the cyber capabilities of US models for GLM-5.3; when Claude's cyber safeguards degraded its attempt on Fable 5, it switched to an older Claude model and another lab's model. In a Moonshot case, requests to Kimi models were secretly forwarded to Claude, including one from a user Anthropic believes is PLA-affiliated who wanted to analyze CCTV footage of a specific person. Why reasoning matters: a pretrained Claude trained on Opus 4.8 transcripts scored 80.4% on SWE-bench Verified with full reasoning and 71.8% with summarized reasoning.

Also from Anthropic's research team: when Claude works around a wall

On October 9 (corrected October 12), Anthropic's alignment team reported four kinds of unintended actions in its evaluations and internal use, including exploiting a basic SQL or command-injection flaw, submitting a sensitive form on a real website, getting around a token or fee gate, and using URL shorteners to dodge a fetch tool's URL limit. In one case Haiku 4.5 submitted an invented tip to the Philadelphia police's homicide tip form; it was flagged as spam. Impact was "minimal," and none touched customer data. Anthropic has turned off live internet access for all internal evaluations until its monitoring is proven. "Most are forms of persistence, in which Claude, when it cannot complete a task as given, works around a restriction instead of stopping."

14:03

Anthropic Caught Chinese Labs Copying Claude.STARTUP HAKK · Oct 14

Gemini 4 Argon is finally for everyone. Two weeks after a defenders-only preview, at $2/$10. #

October 13, 2026

Google made Gemini 4 Argon generally available on Tuesday, October 13: in the Gemini API, in Google Antigravity for AI Ultra subscribers, and in the Gemini app and AI Mode in Search for AI Pro and Ultra subscribers. Issue 26 covered its September 30 reveal to vetted cyber defenders. Introductory API pricing is $2 input / $10 output per million tokens, with cached input 95% off; VentureBeat lists the later rate as $4/$20. The API has a 1M-token unified input-output limit and a new "continuation protocol."

Ars Technica notes the consumer rollout is paid tiers only. Business Insider and The Decoder report Google staff are already testing a stronger Gemini 4 model codenamed "Carbon."

Gemini 4 Argon Is Out (weekly AI news)WorldofAI · Oct 13 · rumor segments excluded here

Meta's double week. A frontier model on Monday, the strongest US open-weight model on Thursday. #

October 12 and October 15, 2026

Meta Superintelligence Labs, led by Alexandr Wang, shipped two very different releases three days apart. Muse Wonder (October 12) is proprietary, multimodal and built for long-running agents, with "Active Memory" and "Active Focus" (it asks permission before consequential actions). Muse Spark Open (October 15) is an 816B-parameter open-weight model that Meta calls "the most capable US open-weight model to date."

Code Review leaves preview, three times faster. And long sessions stop paying to wake up. #

Code Review GA October 15, 2026 · releases 2.1.296–2.1.301, October 9–15

Claude Code Review came out of research preview on October 15 for Team and Enterprise plans. Reviews now finish in 6–10 minutes, down from about 20, and findings are shorter: the problem, what triggers it and the fix "in a few sentences." Admins choose Quick, Standard or Deep per repository. Eligible Enterprise organizations can claim at least $10,000 in Code Review credits by January 15, 2027 (credits expire 30 days after claiming).

Versions 2.1.297 and 2.1.298 were not published.

OpenAI's hardest week was inside its own building. Fired safety researchers, and an employee arrested over a hidden microphone. #

October 9–15, 2026

On Friday, October 9, OpenAI said it had fired three safety researchers, Tomek Korbak, Jasmine Wang and Mikita Balesni, for a "breach of trust" after an investigation found they "violated clear policies on handling sensitive information." Korbak said he was told he was fired over how he communicated with METR, the independent evaluator OpenAI brought in after July's Hugging Face incident; "talking to METR" was his job, he said. Balesni wrote that the three were "fired for prioritizing safety over the near-term interests of OpenAI as a corporation." OpenAI said the firings "were not about safety concerns or speaking out."

The same day, federal agents arrested an OpenAI employee. Prosecutors say Abdulrahman Alfozan, on the product team since July 2025, put a hidden recording device in a USB port behind a TV in an executive conference room, and was caught on a camera OpenAI had placed there. He is charged with intercepting oral communications. OpenAI said it found "no indications" of access to model weights, source code or customer data.

Products kept shipping

• GPT-Live-1.1 (Oct 15) for ChatGPT, Codex and dots: 13 new voices and better French, German, Hindi, Mexican Spanish, Korean and Japanese. Rolling out to all users; API "coming soon."
• Dots (Oct 9): create your always-on dot from the ChatGPT mobile app; dots can now start and continue work in Codex.
• Rosalind Academic Plan (Oct 15): US university labs get up to 30 free Business seats for 12 months, plus the GPT-Rosalind life-sciences model after a trusted-access review.

Who gets to check the labs? Anthropic lets METR inside; Washington keeps its own rules secret. #

October 13 and October 15, 2026

On October 15, Anthropic said it has given a METR team "employee-like access": six METR employees with company hardware, office space and access to internal tools "roughly on par" with Anthropic's own risk-assessment teams. METR will publish its analysis, and Anthropic can't veto publication; it can only require redactions in narrow categories (customer data, privileged legal material, model architectures). Anthropic compares it to the resident examiners regulators place inside big banks.

It's one of three access tiers: SecureBio is reviewing the chem-bio chapter of Anthropic's August Risk Report; Accenture and Faculty are working toward evaluations and red-teaming; early talks include Hugging Face. "We believe every frontier AI company should do the same."

Meanwhile, the government's review stays secret. In Lawfare on October 13, Adam Conner (Center for American Progress) argues the White House's "voluntary framework" under June's Executive Order 14409 has become a de facto licensing system: covered models, set by a classified NSA threshold, go to the government up to 30 days before release. The text was finalized in August but not released; after a Protect Democracy FOIA suit, the government agreed to release it by October 30.

Almost nobody asks Claude who to vote for. Anthropic's midterm numbers, and what changes before November 3. #

October 14, 2026

Anthropic analyzed 250,000 US Claude.ai conversations from September 21–27 with its privacy-preserving method. Politics or social issues came up in about 5.75%, elections or voting in 0.51%, requests for a voting recommendation in about 0.01%, and rule-breaking election requests in about 0.002%.

• A TurboVote banner (Democracy Works) still appears for registration and polling questions; later in October Claude.ai will answer "where do I vote?" with state-specific official information, and on election night it will show official results.
• Opus 5.5, Sonnet 5.5 and Haiku 5.5 score 99% on Anthropic's political evenhandedness evaluation. Outside testers found Claude rarely refuses but sometimes adds unsolicited caveats.
• Red-teamers from ISD and the McCain Institute found Claude could miss a malicious campaign split into harmless-looking pieces. Anthropic added classifiers for multi-step misuse and banned operators of a fake news site producing pro-Russian, Iranian and North Korean midterm coverage.

Voice agents got good enough to answer the phone. Grok Voice 3.0 tops the rankings; GPT-Live gets 13 new voices. #

October 14–15, 2026

On October 14, xAI (now SpaceXAI) released Grok Voice 3.0, #1 on Artificial Analysis's Speech-to-Speech Index at 86.9, ahead of GPT-Live-1 (83.2) and Gemini 3.8 Live (82.6). It scores 97% on Full Duplex Bench (pauses, turn-taking, interruptions) and 71.6% on τ-Voice support calls, close to GPT-Live-1's 74.5% at 40% lower latency. API price: $0.08 per minute.

The retail-relevant number is xAI's own: Starlink call centers use it on 20,000 calls a day, about 70% resolved with no human, and one in four sales calls ends in a purchase (vendor-reported). A day later OpenAI rolled out GPT-Live-1.1 with 13 new voices (story 10).

The money kept coming while the market flinched. Ramp at $60B, Modal at $15.75B, and Character.AI in bankruptcy. #

October 13–15, 2026

Ramp raised $1.85B at a $60B valuation on October 15, four months after raising at $44B. It reports $2B+ annualized revenue and 85,000 customers, and is building agents into procurement and accounting plus a model router for token spend. On October 14, AI cloud Modal raised $750M at $15.75B; Axios reports it passed $1B in annualized run rate.

• Character.AI filed for Chapter 11 on October 13 and is selling some assets to Disney for $20M plus a technology license; it will shut down if no buyer appears.
• Markets wobbled. By Thursday's close the S&P 500 was 1% off its week-ago record and the Nasdaq 3%+ below its peak. NBC points to a Financial Times report that Anthropic's quarterly revenue growth has slowed, plus memory-chip and data-center bond worries.

Shopify opens a billion listings to any AI agent. Payments companies race to own the checkout. #

October 13–15, 2026

On Tuesday, October 13, Shopify said shoppers can let any AI agent browse and buy from its catalog of over a billion listings, not just partners. You connect the agent to your Shop account; checkout happens in the conversation with Shop Pay; you can set a budget cap or shipping rules; and you approve every purchase.

• Stripe + Constructor (Oct 13) launched "Agentic Checkout," combining product discovery and payment in one agent conversation.
• Glance + PayPal (Oct 15): native PayPal checkout inside Glance, an AI shopping app preinstalled on millions of phones (paid press release).

LLM, SLM, SSM: one letter apart, very different ideas. A small language model is about size. A state space model is about how it reads. #

This issue is full of small models: Taalas etched Llama 3.1 8B into a chip (story 03), and Anthropic's self-replication study ran a 9B model on smart-TV boxes (story 05). Two acronyms keep coming up around models like these, and they're easy to mix up. An SLM (small language model) is an ordinary chatbot model shrunk to run on a laptop or phone. An SSM (state space model) is a different design: instead of rereading everything you've said, it carries a fixed-size running summary. They can overlap: some of 2026's best small models are SSM hybrids.

Size versus architecture

A transformer's core trick is attention: every new word it writes is compared with every earlier word in the conversation, which the model keeps in memory (the "KV cache"). That's why transformers are so good at pulling one exact detail from page 40, and also why long chats get expensive: memory grows with every token, and the work grows roughly with the square of the context length.

An SLM doesn't change any of that; it just has fewer, smaller layers, so it is cheaper, faster and private enough to run offline. The cost is less knowledge and weaker reasoning on hard problems. An SSM changes the mechanism. Each new token updates a compact state and moves on, so the work grows linearly and memory per conversation stays flat, however long the input. The cost: a fixed-size summary has to forget something, and pure SSMs are worse at exact recall ("what was the phone number on page 40?"). So the SSMs that shipped in production are hybrids: mostly Mamba layers, plus a few attention layers for precise lookup.

Small language models you can run today

• Google Gemma 4 (April 2, 2026): the E2B and E4B versions are built for phones, take text, images and audio, have a 128K context window and run fully offline in Google's AI Edge Gallery app. Apache 2.0.
• Microsoft Phi-4 family: Phi-4-mini (3.8B, 128K context, function calling), Phi-4-multimodal (5.6B), Phi-4 (14B) and Phi-4-reasoning-vision-15B (March 2026), all MIT-licensed.
• Apple's on-device model: the roughly 3B-parameter model behind Apple Intelligence, which app developers can call through Apple's Foundation Models framework, with no cloud and no per-token bill.
• Alibaba Qwen3 (0.6B to 8B) and Meta Llama 3.1 8B: the open workhorses that hobbyists and chip makers build on. Llama 3.1 8B is the model Taalas hard-wired into ChatJimmy.

State space models in production

• Mamba (Albert Gu and Tri Dao, December 2023) made SSMs competitive with transformers. Mamba-3 (March 2026, ICLR 2026) matches Mamba-2's quality with half the state size.
• IBM Granite 4.0: hybrid models with roughly nine Mamba-2 layers for every attention layer; IBM says they cut memory use by over 70% versus transformer-only models. Small sizes (Micro, H-Tiny) are SLMs and SSMs at once. Apache 2.0.
• NVIDIA Nemotron 3 Nano (December 2025): a hybrid Mamba-Transformer mixture of experts, 31.6B total, 3.2B active per token, with a 1M-token context and up to 3.3x the throughput of similar open models (NVIDIA-reported).
• AI21 Jamba: the first production-grade hybrid, stacking Mamba and attention layers to handle 256K-token inputs. Mistral's Codestral Mamba (7B) showed a pure SSM could write code.

Cincinnati and Ohio AI. Kroger ties corporate jobs to AI, and what's on next. #

Kroger will cut 3,000 corporate jobs over three years as it leans into AI

Memo October 12 · reported October 13, 2026

Cincinnati-based Kroger told employees it will lay off 3,000 corporate employees over the next three years as it ramps up AI. CEO Greg Foran, the former head of Walmart US, wrote: "Technology is reshaping what's possible… We need to adopt capabilities, reduce duplicative work and look closely at where work is best fit to be done." Kroger has 409,000 employees and 2,700+ stores; shares fell 2.2% on Tuesday. It is the region's largest employer publicly tying corporate headcount to AI; expect other retailers' cost plans to be framed the same way.

Miami University leans in, carefully

Axios Cincinnati (October 13) profiled Miami University's approach: a new Miami Center for AI, approved in September, and the "AI in the Majors" plan to bring AI into every department's curriculum by 2027–28.

Upcoming events

From the editor · CincyAI

I'm a CincyAI Champion. Come join us.

CincyAI for Humans, founded in October 2023, is Ohio's largest AI community, and its monthly meetups at UC Digital Futures stay free. Champions is the paid members' portal that keeps the conversation going between meetups:

• Human Huddle interest groups and Champion Matchups with peers
• Monthly office hours with co-hosts Kendra Ramirez and Helen Todd
• A member directory, a Cincinnati-area AI job board and the events calendar

It starts with a 30-day free trial, and the annual plan is $50 off. Become a CincyAI Champion · cincyai.org

Twelve tips we haven't run before. Each one uses something shipped or documented this week. #

Every tip was checked against past issues of The AI Enthusiast; none has run here before. Tips 1 to 4 are for everyone, 5 to 10 for Claude Code, and 11 and 12 for anyone building agents. The REA and Jev tips are in stories 01 and 04.

5:11

New Claude Motion is InsaneThe AI Advantage · Oct 14 · quickest overview

17:06

Claude NEW Motion is Insane (8 Real Use Cases)Ben AI · Oct 15 · best for ideas