Check out the resources I offer below and sign up for my new newsletter!
Share
mAIn Street #296: Nvidia prices are going up; Schools stumble using AI; Apple cuts AI-related jobs; Getting started with GrokBot and other agents
Published 3 days ago • 11 min read
Nvidia server price hikes, Kaiser triage concerns and OpenAI’s safety-law reversal lead Monday’s AI news.
Monday, August 24, 2026
On September 9, I’m hosting ChatGPT for Marketing: Using AI So It Doesn’t Look Like AI, a 90-minute workshop for marketers and business professionals who are tired of generic copy and look-alike visuals. We’ll work with writing, images, fonts and web pages, then use real examples to show how references, clear instructions, repeated edits and human judgment produce work that feels specific to you. See the workshop details and register.
Some of Nvidia’s largest customers have reportedly been told that servers using Vera Rubin and Grace Blackwell chips will cost more when shipments begin next year. Reuters could not independently verify the report, and Nvidia had not commented, but rising memory costs are becoming a direct pressure on AI budgets.
More than a dozen clinicians, therapists and experts in California and Wisconsin described weeks-long delays, missed histories of self-harm and patients routed to lower levels of care. Kaiser says licensed clinicians make treatment decisions and its technology supports their judgment. The dispute shows why health-care automation needs visible human ownership.
New records show Farnsley Middle School designed the booklet and Jefferson County Public Schools printed it. The agenda mislabeled states and distorted science graphics, a public reminder that generated material still needs an accountable human review before money is spent or students see it.
About 100 roles are reportedly coming from Vision Pro, with another 100-plus affecting Siri and the Intelligent Systems Experience group. Apple confirmed the cuts and said the reshuffle will also create new roles. AI investment can move jobs around inside a company even while total spending grows.
The company wants SB 53 to require monitoring of frontier models during training and evaluation, along with stronger cybersecurity through the development cycle. OpenAI opposed the law before it passed. It now says state protections can form a national baseline while Congress remains stalled.
A three-level capability map tells an AI agent where its job ends.
On Friday, SpaceXAI widened access to Grok Bot, an early-beta agent that runs on a persistent cloud computer, works through signed-in apps and websites, and keeps going after a user steps away. The company says it is now included with SuperGrok Plus, Cursor Pro+, and all Cursor Teams plans.
Grok Bot joins ChatGPT Work, Claude Cowork and Gemini Spark in a growing class of products that can retrieve material, use connected tools and complete multi-step jobs. That makes delegation design a basic workplace skill. Before an agent gets access to an inbox, drive or customer system, map the job into three levels: safe to delegate, delegate with review, and human-owned. This is capability mapping.
SpaceXAI expanded Grok Bot access on Aug. 21. Logo source: Wikimedia Commons.
Start with the decision the work must support
Suppose you oversee a software subscription used by 40 employees. The contract expires in 30 days. Pricing notices sit in email, usage data lives in a spreadsheet, support history sits in a help desk, and legal terms are stored in Drive. You need a defensible recommendation: renew, renegotiate or replace.
The model can interpret the material. The surrounding product supplies file search, app access, browser use, artifact creation or scheduling. Your organization supplies permissions, policy, expertise and accountability. A capability map separates those responsibilities before the work begins.
Capability mapping turns a complete job into clearly bounded steps. Photo: Vitaly Gariev via Unsplash.
Level
What belongs there
Safe to delegate
Retrieve the current contract, pricing notice, usage figures and support history from approved locations. Normalize dates, fees and license counts. Remove duplicates. Build an evidence table that links every figure to its source and flags missing or conflicting information.
Delegate with review
Classify support problems by business impact. Calculate cost per active user. Identify underused licenses. Compare shortlisted alternatives using official product, pricing and security sources. Draft the negotiation memo for a person to check.
Human-owned
Decide how much disruption, legal exposure and security risk the company can accept. Approve vendor contact, purchases, cancellations and data changes. Accept terms and sign the contract.
These boundaries match the approval controls described in Grok Bot’s own documentation, which calls for explicit approval around messages, purchases, deletion, permission changes and legal terms.
Choose the product by where the job happens
ChatGPT Work can research across connected apps and files and produce finished documents, spreadsheets, presentations, reports and Sites. Claude connectors let Cowork retrieve data and take supported actions under the user’s existing source permissions. Gemini Spark can use connected apps, skills and signed-in websites, although Google currently limits Spark to personal Google accounts with an AI Pro or Ultra subscription. Grok Bot’s cloud computer can work through browser-only systems, subject to its beta limitations and eligible plans.
Ask one practical question: does this product surface have approved access to the sources and actions the job requires? A capable model confined to a chat window may still need you to ferry every file and result. An agent with the right access can complete more of the job. Begin with read-only access, then widen its authority only after repeated, accurate results.
Give the agent a real assignment
Job: Prepare the [vendor] renewal decision for [decision-maker] by [date].
Outcome: Produce a recommendation supported by current costs, usage, service history, contract terms, switching costs and unresolved questions.
Approved access: Use only [Drive folder], [inbox label], [usage spreadsheet] and [help-desk queue]. Keep all access read-only. Do not open HR, payroll or unrelated customer records.
Work: Find the latest source for each item, remove duplicates, reconcile dates and totals, calculate cost per active user, and compare the approved alternatives using their official pricing and security pages.
Deliverable: Create a two-page decision memo and a source table. Put the recommendation first. Separate facts, assumptions, missing information and questions for the vendor.
Quality checks: Cite every number and contract claim. Use the current plan name and price. Show conflicting terms side by side. Never fill a gap with a guess.
Stop and ask: Pause if a source is inaccessible, important terms conflict, or the next step would send a message, change a record, reveal credentials, spend money, cancel service or accept terms. Return the proposed action for approval.
Save this brief as a skill, routine, workspace agent or scheduled task in the product you use. Give the agent approved connections so it can retrieve the material itself. A one-time file upload remains the fallback when a connector is unavailable or company policy requires manual handling.
For managers and owners
Run one job as a small pilot. Track source coverage, corrections, completion time and approval requests. Three clean runs within the same read-only scope can justify saving the process for reuse. Missing sources, unclear ownership or frequent corrections should shrink the scope before the agent receives more access.
Tomorrow’s move: Choose one recurring job due this week and spend 15 minutes building its three-level map. Give the agent only the first-level work. Review the result against your normal standard, then decide whether one second-level step has earned a supervised trial.
AI Tool Spotlight
Claude Academy puts Anthropic’s practical training in one place
The free learning hub covers Claude’s main products, workplace use and a shared framework for responsible delegation.
Claude Academy gathers Anthropic’s official learning materials for Claude.ai, Cowork, Claude Code, Claude Tag and the Claude API. The courses cover drafting, data analysis, projects, artifacts and tool connections.
Where it earns its keep
New users: follow product-specific lessons instead of stitching together advice from old tutorials.
Teams: use the AI Fluency track to establish shared language around delegation, description, discernment and diligence.
Builders: move from everyday Claude use into Claude Code and API training without changing learning hubs.
Know before you start
The material is specific to Anthropic products. It is a useful starting point for Claude users, while company policy, data permissions and human review still determine what is appropriate at work.
The increases would affect systems shipped early next year, with the final price tied to chip generation and memory configuration. Higher infrastructure costs can flow into cloud rates and business AI budgets.
Roughly 100 roles reportedly come from Vision Pro, with another 100-plus affecting Siri and the Intelligent Systems Experience group. Apple confirmed the cuts and said the change will also create new roles.
The startup says it has 600 customers and reached a $1 billion valuation amid a U.S. accountant shortage. Its audit feature lets human accountants inspect the numbers and calculations chosen by AI agents.
A custom setup lifted Claude Opus 5 from a 30% score to 100% on the ARC-AGI-3 interactive benchmark. Strong tools, memory and oversight can produce large performance gains.
AI-related corporate debt has climbed from $12.5 billion in 2025, and recent deals have required larger concessions to attract buyers. The buildout remains well funded, though money is getting more expensive.
The booklet mislabeled states and distorted a periodic table and moon phases. The district called it an isolated incident and stressed the need to review material before distribution.
OpenAI wants SB 53 to add monitoring during model training and evaluation and stronger cybersecurity through development. The company opposed the law before it passed and now sees state protections as a possible national baseline.
The company cut its default data-retention period from 30 days to seven and now requires a case code for access. Both settings can be overridden, which keeps attention on how local agencies use the system.
The eight-week Foundry program includes weekly live sessions, while HeyGen avatars provide individual feedback during practice pitches and board meetings. Students get more rehearsal time, and instructors lend out a digital version of their judgment.
Thirty-seven states have offered AI guidance for schools, and districts in places such as Charleston and Utah are training teachers and students. Verification, privacy and knowing when to avoid a chatbot are becoming part of digital literacy.
Workers described weeks-long delays and cases in which histories of self-harm were missed. Kaiser says licensed clinicians make clinical decisions and its technology supports their judgment.
The team trained a model on roughly 500,000 experimental data points to identify the initiator sequence, which appears in about 60% of human genes. The result may help researchers find mutations linked to diseases such as cancer.
The startup keeps donated tissue alive for up to a month, tests compounds predicted by its model and feeds the results back into the next round. It says the loop now produces a candidate about every six weeks.
Researchers analyzed blood from more than 4,000 people and used AI to separate likely strong responders from weak ones. The work is an early research result, not a clinical test, but it points toward more personalized vaccination plans.
The complaint challenges the company’s sleep-stage claims for rings that cost $300 or more. Oura disputes the allegations, cites independent research and says its ring is not a substitute for a clinical sleep study.
Higgsfield later confirmed that videos from Matti Haapoja and Sam Kolder were paid collaborations. The reaction shows how quickly creator trust can erode when AI endorsements and commercial ties are unclear.
The 20-minute film is scheduled to premiere at a Long Island festival in September, and Joel is not involved in the production. The project places cultural preservation and consent over a person’s likeness in the same frame.
The control sits in Accounts Center under Activity from other businesses. It limits off-platform personalization, though it does not stop Instagram from using activity inside the app or settle separate model-training choices.
The IAB cites an NYU Stern study while recommending disclosure when AI use could materially mislead consumers. The finding gives marketers a concrete measure of the trust cost that can accompany generated creative.
Recent rulings have treated some training as fair use while penalizing pirated source libraries or systems built to compete directly with the source. Businesses using generated content still face many unsettled legal questions.
The London lab says Faraday reproduced findings from published papers using a 27-billion-parameter model. The company’s claim points to how training methods and agent design can stretch smaller models.
Robots at the Beijing event beat human sprint records, then faced less glamorous tests in warehouses, restaurants and charging stations. Commercial value will depend on reliable work in imperfect real-world settings.
The offering is set to become the largest primary follow-on sale by a Hong Kong-listed company. Alibaba plans to put the proceeds into its full AI stack as heavy capital spending weighs on current profit.
About 20 AI, data and cybersecurity specialists will focus on intelligence support, critical infrastructure and regional monitoring. The effort brings military AI development into a standing two-country unit.
Clients are demanding more output for less money, and smaller firms are winning work with faster pilots and flexible terms. The shift is weakening the old model that tied revenue growth to larger teams of junior coders.
mAIn Street gives nontechnical readers useful AI news they can put to work.
AI is moving deeper into medical care, professional work, personal finance and the infrastructure debate. Wednesday, August 26, 2026 On September 9, I’m hosting ChatGPT for Marketing: Using AI So It Doesn’t Look Like AI, a 90-minute workshop for marketers and business professionals who are tired of generic copy and look-alike visuals. We’ll work with writing, images, fonts and web pages, then use real examples to show how references, clear instructions, repeated edits and human judgment...
AI pay packages, school rollout caution and supervised clinical agents lead Tuesday’s AI news. Tuesday, August 25, 2026 On September 9, I’m hosting ChatGPT for Marketing: Using AI So It Doesn’t Look Like AI, a 90-minute workshop for marketers and business professionals who are tired of generic copy and look-alike visuals. We’ll work with writing, images, fonts and web pages, then use real examples to show how references, clear instructions, repeated edits and human judgment produce work that...
Micron plans a $10 billion AI lab in Boise, data-center demand reaches U.S. factories, and Google puts Gemini inside Chat. Friday, August 21, 2026 On September 9, I’m hosting ChatGPT for Marketing: Using AI So It Doesn’t Look Like AI, a 90-minute workshop for marketers and business professionals who are tired of generic copy and look-alike visuals. We’ll work with writing, images, fonts and web pages, then use real examples to show how references, clear instructions, repeated edits and human...
