Ask The PhD Community

HigherEd AI Daily

May 14 – Metrics, Models, and the Devices Reshaping Campus AI

Thursday, May 14, 2026

Today's signals point to a shared question for higher education; as AI embeds deeper into research and operations, are we measuring the right things, choosing the right platforms, and preparing for the devices students will actually carry into class?

The Rundown AI — Governance

Amazon's AI Scoreboard Is Warping How People Work

Amazon set an internal target for 80 percent of its developers to use AI weekly and began ranking staff by model and token consumption. Employees told the Financial Times the result has been predictable; workers are now running the internal MeshClaw agent on unnecessary tasks just to lift their numbers. The company says token counts are not used in performance reviews, yet it has already pulled back visibility of the metric to individual staff and their managers.

The pattern is being called "tokenmaxxing," and reporting suggests Meta and other large employers are wrestling with similar dynamics. The lesson is older than AI; when an institution rewards a proxy, people optimize for the proxy and not the underlying outcome.

Why it matters for campuses

Campus leaders under pressure to demonstrate "AI adoption" should resist building dashboards around usage counts. License seats activated, prompts logged, or tools opened are activity, not learning or research quality. Faculty senates, provosts, and CIOs designing AI strategy should anchor evaluation in outcomes that matter: improved research throughput, time recovered for student-facing work, evidence of better teaching artifacts. Measure what you actually want to grow.

Read More

(subscription may be required)

The Rundown AI — Policy

Anthropic Takes the Enterprise Lead From OpenAI

For the first time, Ramp's monthly AI Index shows Anthropic ahead of OpenAI in paid business adoption. Anthropic climbed to 34.4 percent of tracked adoption in April while OpenAI dipped to 32.3 percent, capping a roughly fourfold rise in Anthropic usage since 2025. Ramp draws its signal from corporate card and invoice data across more than 50,000 U.S. businesses, so it is a spend indicator rather than a full market share view.

Most of the swing traces to Claude Code, which has pulled Anthropic out of developer-only territory and into finance, legal, and research workflows. The report also flags real risks, including recent Claude reliability incidents and cost pressures relative to OpenAI and open-source alternatives.

Why it matters for campuses

Institutions standardizing on a single AI vendor for the next academic year should treat the market as unsettled. Provosts and CIOs negotiating enterprise contracts can use this volatility as leverage; insist on portability, exit terms, and the ability to run two providers in parallel. The signal for instructional designers is more practical; Claude's growth in research and analytical workflows means more students and faculty will arrive on campus already fluent in it, not just in ChatGPT.

Read More

The Rundown AI — Access

Google Unveils Gemini-Native Googlebooks for Students and Faculty

Ahead of next week's I/O event, Google introduced a new line of AI-native laptops called Googlebooks, built in partnership with Dell, HP, Lenovo, Acer, and Asus, alongside a cross-device layer called Gemini Intelligence. The devices ship this fall and blend ChromeOS, Android, Google Play, and Gemini in one stack, with a "Magic Pointer" cursor that triggers full-screen Gemini interactions and the ability to run Android apps and files natively.

Other releases include a Create My Widget tool, a Rambler dictation tool that removes filler words, and on-device Gemini auto-browse inside Chrome. Where Apple's Siri revival is still pending, Google is now weaving its assistant directly into the operating system rather than layering it on top.

Why it matters for campuses

If Chromebook-style devices become AI-native by default, the calculus for student device programs changes. Equity-focused campuses with large Chromebook deployments, community colleges, and K-12 partners feeding into your institution should start asking now what an "AI-included" device baseline means for academic integrity policy, accessibility, and bring-your-own-device guidance. Decisions you make this summer will define what fall 2026 freshmen walk into.

Read More

The Rundown AI — Research

Adaption's AutoScientist Automates Model Fine-Tuning

Adaption, the startup founded by former Cohere VP of Research Sara Hooker, introduced AutoScientist, a system that automatically customizes AI models for specific tasks by iterating on both the training data and the training procedure. In internal tests, AutoScientist outperformed Adaption's own expert-tuned models by an average of 35 percent and lifted success rates from 48 percent to 64 percent across eight industries, including legal, medical, and financial domains.

Only a small number of researchers worldwide know how to fine-tune a frontier model well, and most of them work at a handful of labs. Tools that automate that expertise lower a barrier that has kept domain-specific models out of reach for organizations without elite ML teams.

Why it matters for campuses

Universities sit on highly specialized corpora; clinical notes, historical archives, discipline-specific journal collections, departmental teaching materials. If automated tuning matures, mid-sized research universities and even individual labs may be able to produce credible domain models without recruiting a frontier ML team. Research offices and libraries should begin scoping which datasets they would tune on, what data-use and IRB constraints apply, and who owns the resulting model weights.

Read More

Tool of the Day

Transformers in Practice (DeepLearning.AI)

DeepLearning.AI just launched a new course, taught by Sharon Zhou, VP of Engineering and AI at AMD, that opens up the inside of transformer-based models. It covers how generation works token by token, what attention and positional encoding actually do, why hallucinations happen, and how techniques like RAG, KV caching, and speculative decoding affect inference. It is built for practitioners and informed users, not pure researchers, which makes it well suited to faculty, instructional designers, and graduate students who need a working mental model rather than full ML theory.

Try it: enroll one faculty member or instructional designer this week, then ask them to convert two of the course's interactive visualizations into a 20-minute "how LLMs actually work" segment for your next AI literacy session with faculty or students.

Visit DeepLearning.AI

[CLOSING PLACEHOLDER: Dr. Green, please replace this bracketed line with your 2-sentence closing in your own voice before sending. This automated run could not request it interactively.]

Dr. Ali Green

Sources for This Edition

The Rundown AI (rundown.ai)
Financial Times (ft.com)
Ramp AI Index (ramp.com)
Google Blog (blog.google)
Adaption Labs (adaptionlabs.ai)
DeepLearning.AI / The Batch (deeplearning.ai)

askthephd.com
 | 
askthephd.substack.com
 | 
Unsubscribe

HigherEd AI Daily; Curated by Dr. Ali Green