Welcome back everyone! 👋

This week feels like a turning point. The tools are getting materially better, enterprise adoption is accelerating, and the gap is widening between people who experiment and people who build real workflows.

In AI Education, we continue the large language model series with Part 3, walking through what actually happens after you hit enter, from tokens and context windows to attention and next token prediction.

This week’s 10-Minute Win is a Savings Goal Sprint Planner that turns “I should save more” into a concrete target, timeline, and calendar nudges you will actually follow.

And in Founder’s Corner, I share three bold predictions for how AI will reshape the workplace in 2026, plus what you can do now to stay ahead instead of scrambling later

1) OpenAI releases ‘code red’ GPT-5.2 update to ChatGPT

Summary: OpenAI rolled out GPT-5.2 as its “most capable model series yet for professional knowledge work,” framed internally as a “code red” response to Google’s Gemini push. The new Instant / Thinking / Pro models are tuned for real work: better spreadsheets, presentations, coding, image understanding, long-context reasoning, and tool use, with higher accuracy and fewer hallucinations than 5.1

Why it matters: Your default AI coworker just got a material upgrade in the exact areas that move money—analysis, modeling, and long-running projects. If you’re still treating AI as a toy, this release makes that a bad strategy.

2) Disney making $1 billion investment in OpenAI, will allow characters on Sora AI video generator

Summary: Disney is putting $1B of equity into OpenAI and signing a three-year licensing deal that lets Sora and ChatGPT Images use 200+ Disney, Pixar, Marvel, and Star Wars characters in user-generated content. Some Sora clips will stream on Disney+, and Disney will roll ChatGPT and OpenAI APIs into internal workflows.

Why it matters: This is Hollywood’s biggest embrace of gen-AI so far. Disney isn’t just “experimenting” — it’s turning AI into a new distribution and monetization layer for its IP while trying to keep tight control over brand and guardrails.

3) The state of enterprise AI – 2025 Report

Summary: OpenAI’s new report shows enterprise AI is not dabbling anymore. ChatGPT now serves 7M+ workplace seats, Enterprise seats are up ~9× YoY, and weekly Enterprise messages are up ~8× since late 2024. Workers report 40–60 minutes saved per active day (heavy users >10 hours/week), and 75% say AI improves speed or quality. Custom GPTs and Projects are exploding (19× growth in weekly users), with ~20% of Enterprise messages now flowing through these persistent, workflow-specific agents.

Why it matters: The gap isn’t “who has AI” anymore — it’s who has it wired into core workflows. If you’re not building your own small stack of agents and automations, you’re on the wrong side of that divide.

4) Trump approves sale of more advanced Nvidia computer chips used in AI to China

Summary: The U.S. will allow Nvidia to sell its H200 AI chips to “approved customers” in China, reversing tighter export limits. The H200 isn’t Nvidia’s top Blackwell/Rubin tier but is still a major step up from what China could buy before. Nvidia applauded the move; a bloc of Democratic senators warned it hands China “transformational” AI capabilities with military and cyber implications.

Why it matters: Export controls are one of the few real brakes on the global AI arms race. Loosening them boosts Nvidia and short-term U.S. business interests, but it also narrows the compute gap with China. That trade-off will echo through both geopolitics and AI competition.

5) Five debt hotspots in the AI data centre boom

Summary: Reuters walks through how the AI data-center buildout is being financed with increasingly complex debt: ~$75B in recent investment-grade issuance from Big Tech, multi-billion “off-balance-sheet” project structures (e.g., Meta–Blue Owl, Oracle deals), a spike in junk bonds from AI infra players, a growing role for private credit, and new asset-backed securities tied to data-center rents. Central banks, including the Bank of England, are already flagging pockets of risk.

Why it matters: AI isn’t just an equity story; it’s a leveraged credit story. Great on the way up, ugly if expectations reset. If you care about macro and markets, you need to track the plumbing behind the AI hype.

So far you have:

• A mental model of what an LLM is: a text pattern learner
• A training story: how it learns from data and adjusts billions of settings

Now we zoom into the moment that matters most to you:

You type a message. What happens next?

Step 1: Your words become tokens

The model does not work directly on raw characters or full sentences. It works on tokens.

A token is a small piece of text. When you write:

“Help me understand why my grocery spending went up this month.”

The system turns that sentence into a sequence of tokens chosen from the model’s vocabulary.

The exact split does not matter to you. What matters is:

• The model sees your input as a sequence of tokens
• It treats earlier messages and system instructions as tokens too

Everything it reads is in that token form.

Step 2: Tokens must fit in the context window

The model has a context window, which you can think of as a page with a strict size limit.

• The window holds a fixed number of tokens
• Your current prompt, earlier chat history, and any pasted data all share that space
• Anything that does not fit is simply not visible to the model

If you paste in:

• A short bank statement and a clear question → likely fits
• Years of raw transaction history → most will not fit

This is why structure and focus matter. You want the right tokens on the page.

Step 3: The model decides what to focus on

Inside the model there is a mechanism often described as attention.

In plain English, attention is:

The way the model looks across all tokens in the context window and decides which ones are most important for predicting the next token.

Analogy: You get a long email from your bank. Before you reply, you reread the few sentences that matter and skim the boilerplate. You do not weigh all lines equally.

The model does something similar in math:

• It looks at all tokens
• It gives some tokens more weight than others
• Those weighted tokens influence what it predicts next

So if you ask about grocery spending and include a short summary plus a few labeled sections, the model is more likely to focus on those high-signal pieces.

Step 4: Predicting one token at a time

Even during a long answer, the core behavior is simple:

1. The model looks at the current sequence of tokens in the context window
2. It produces a probability for each possible next token
3. It picks one token, often with a bit of randomness
4. It adds that token to the sequence
5. It repeats the process, token by token, until it stops

From your point of view, it looks like fluent paragraphs. Under the hood, it is next-token prediction over and over, guided by the patterns learned during training and the context you provided.

Step 5: Why prompt structure matters so much

Because the model:

• Can only see what fits in the context window
• Pays more attention to some tokens than others
• Predicts based on patterns and context

Your prompt design and input structure matter a lot.

Tips:

• Start with a clear goal
• Provide only the key facts the model needs
• Organize long data into chunks with titles and short summaries.

You are making it easier for the model to find the right tokens and patterns inside the window.

Reader questions

Q: If it only predicts one token at a time, why does it sound so coherent?

A: Because it was trained on huge amounts of connected text. It has seen how sentences, paragraphs, and arguments are usually built. Next-token prediction over many steps, guided by those patterns, can produce answers that feel smooth and logical.

Q: Why does it sometimes forget something I said earlier?

A: Often because that part of the conversation has been pushed out of the context window by newer messages. Or because you changed instructions later, and the model now weighs the more recent text more heavily.

Closing this week

At this point you have both sides of the core loop:

• Training time: the model learns patterns by predicting tokens and adjusting its settings
• Chat time: the model turns your text into tokens, fits them into a window, focuses on key pieces, and predicts new tokens

Next week we will focus on the parts that can bite you:

• Hallucinations and confident wrong answers
• Shallow, generic responses
• Bias and gaps from training data
• How to work with these limits instead of being surprised by them

🎯 Savings-Goal Sprint Planner

Why this matters: “I should save more” is not a plan; it’s background guilt. The gap between that feeling and real progress is simple: one specific goal, a clear dollar target, a deadline, and a behavior pattern that supports it. In this 10-minute workflow, you’ll let ChatGPT build your savings planner table for you, then turn one fuzzy goal into a SMART savings sprint with monthly and weekly targets and calendar nudges.

Step 1 — Have ChatGPT build your planner table (2–3 minutes)

Open ChatGPT Free and paste this prompt:

You are a Google Sheets template builder. I want to create a simple “Savings-Goal Sprint Planner” in Google Sheets.

Task:

• Create a CSV-style table starting at row 1 with these columns:
  • Goal Name
  • Target Amount $
  • Target Date
  • Months Left
  • Monthly Target $
  • Weekly Target $
  • Status
• Row 1 should be the headers.
• Row 2 should be example data with formulas where appropriate:
  • Goal Name: Sample – $1,200 Emergency Buffer
  • Target Amount $: 1200
  • Target Date: a date 6 months from today in YYYY-MM-DD format.
  • Months Left: =DATEDIF(TODAY(),C2,"M")
  • Monthly Target $: =ROUND(B2/D2,0)
  • Weekly Target $: =ROUND(E2/4,0)
  • Status: Not started
• Output only the table in plain text, comma-separated, with no explanations before or after.

Goal: I want to be able to copy this output and paste it directly into cell A1 of a blank Google Sheet so the headers and formulas work.

Then:

1. Copy the CSV-style output from ChatGPT.
2. Open a blank Google Sheet, click cell A1, and paste.
3. You should see headers in row 1 and a sample row with formulas in Months Left, Monthly Target, and Weekly Target.

You now have a working planner. You’ll overwrite the sample row in a minute.

Step 2 — Turn your idea into a SMART savings sprint (3 minutes)

In a new ChatGPT message (same chat is fine), paste and fill in:

You are my Savings-Goal Sprint Planner. I want to create one 90–180 day savings sprint.

My rough idea:

• What I’m trying to save for (e.g., trip, emergency buffer, debt payoff, down payment): ___
• Rough dollar amount I think I need: ___
• When I’d like to have it by (month/year): ___
• Approximate monthly free cash I could redirect if I try (even a guess): ___

Tasks:

1. Turn this into a single SMART savings goal (Specific, Measurable, Achievable, Relevant, Time-bound).
2. Suggest a realistic Sprint Length in months (between 3 and 6).
3. Calculate the Monthly Target and Weekly Target amounts needed to hit the goal within the sprint, given that sprint length.
4. List 3–5 behavior rules that will support this sprint (where the money will come from, small cuts, or temporary rules).

Output format:

• SMART Goal (1–2 sentences)
• Sprint Length: X months
• Target Amount: $X by [date]
• Monthly Target: $X
• Weekly Target: $X
• Behavior Rules: bullet list

Rules: If my free cash estimate is clearly too low for the goal in 3–6 months, say so and suggest either a smaller target or a longer sprint.

You’ll get a standardized block you can plug straight into your planner.

Step 3 — Replace the sample row with your real sprint (3 minutes)

Back in Google Sheets, in row 2:

• A2 (Goal Name): short name from your SMART goal (e.g., “Emergency Buffer” or “Summer Trip”).
• B2 (Target Amount $): the Target Amount ChatGPT gave you.
• C2 (Target Date): the date from the SMART goal (or your final choice).

Confirm:

• D2 (Months Left) uses =DATEDIF(TODAY(),C2,"M").
• E2 (Monthly Target $) uses =ROUND(B2/D2,0) or you can manually type the “Monthly Target” ChatGPT gave you if you prefer its sprint length.
• F2 (Weekly Target $) uses =ROUND(E2/4,0).
• G2 (Status): set to In progress.

Under the table, paste the SMART Goal and Behavior Rules so you see them every time you open the sheet.

Step 4 — Add calendar nudges so you actually do it (2 minutes)

Open Google Calendar:

1. Create an event on your next payday titled: Move $[Monthly Target or Weekly Target] → [Goal Name].
   • Set Repeat: monthly (or weekly if you’re saving weekly).
   • In the description, paste your SMART goal and a link to the Google Sheet.
2. Create a second event 30 days from today titled: Savings Sprint Review — [Goal Name].
   • Set it to repeat every month.
   • In the description, add:
     • “Current saved amount: _____”
     • “Status: On track / Behind / Ahead”

This is your built-in feedback loop: each month you update the status, see the math, and either stay the course or adjust.

The Payoff

In 10 minutes, you’ve gone from a vague intention to a concrete, time-bound savings sprint: a planner table that AI built for you, a SMART goal with real numbers and a real date, and recurring reminders that pull the goal back into your field of view. You can reuse the same sheet for future sprints—just add a new row per goal.

Transparency & Notes for Readers

• All tools are free: ChatGPT Free, Google Sheets, Google Calendar.
• Math is simple: targets are based on basic division; tweak sprint length or goal size if the numbers feel unrealistic.
• Behavior is the real engine: the calendar reminders and behavior rules matter more than the formulas being perfect.
• Educational workflow — not financial advice.

I use AI in two distinct worlds: building Neural Gains Weekly and working in my corporate job in the AI and digital technology space. Throughout 2025, it has been easier to experiment, learn, and build with AI in my personal life, and I think that rings true for many of us. PwC’s latest global workforce survey found that 54% of workers have used AI for their jobs in the past year, but only 14% are using generative AI daily, which tells me most people are still in “light experimentation” mode at work. Inside most companies, 2025 has been about getting the right infrastructure in place and running pilots and proof-of-concepts.

2026 feels like an inflection point. The signals are everywhere: corporate restructures to better organize around AI, infrastructure buildouts that lower the cost to deploy AI systems, and a surge in AI agents quietly showing up inside the tools people already use at work. Change is on the horizon, and I want to share three hot predictions for how AI will reshape the workplace, and how you can get ready.

AI Usage Becomes Mandatory, Not Optional

Companies are investing heavily in infrastructure, data systems, and integrations that will accelerate AI adoption. That spending will not stay invisible for long. As those foundations solidify, expectations will change, regardless of our role or level. We will all have to adapt and use AI to stay ahead of the curve and, frankly, to stay relevant in the future of work.

You can already see the early signs. AI skills are starting to show up in job descriptions, and I expect that trend to accelerate in 2026. I think we will start to see AI portfolios being requested in interviews and promotion conversations, with real examples of how you used AI to solve a problem, improve a workflow, or save time for your team. Goals tied to AI usage will quietly make their way into performance reviews, holding employees accountable for growing their skill set to match where their company is heading. The shift will not happen overnight, but the direction is clear: using AI at work moves from “nice to have” to “part of the job.”

AI Won’t Take All the Jobs, But It Will Rewrite the Job Description

AI automation and augmentation will be major themes in the next phase of work, but the real question is how these catalysts will actually impact our jobs. According to the WEF Future of Jobs report, 39% of core worker skills are expected to change by 2030, and AI is expected to replace 9 million roles and create 11 million roles over that same period. This signals a shift away from purely human work toward systems that integrate human and machine intelligence.

AI is going to force human workers to evolve their skills to match what employers need from their workforce. The immediate risk is not “AI is going to replace me in 2026.” The real risk is “Someone who understands how to use AI to clear the easy work will have more time to do the higher-value work than I do.” New roles will emerge and new industries will be born. That is a good thing, and it is not something that can be stopped. This evolution will reward the people who use AI to clear the busywork so they can spend more time solving problems, thinking creatively, and strategizing for the future.

Non-Technical Builders Become a Force

It seems like 2025 was the year of “anyone can code an app,” but that trend has mostly stayed on the personal side. Anyone can open Replit, Google AI Studio, Claude, or ChatGPT and “vibecode” an app or generate code for a specific use case. But we have not seen the same speed and adoption of these types of tools in the workplace. There are AI coding agents that technical employees use in their daily tasks, but the tools for non-developers are lagging. That changes in 2026.

According to Gartner, by 2029, 50% of knowledge workers will develop new skills to work with, govern, or create AI agents on demand. We are still in the early stages, but in 2026 more companies will begin launching low code and no code tools that allow non-technical employees to build prototypes that would have taken months in a typical IT project lifecycle. Prompt and context engineering through natural language will replace the need to write most of the code and will accelerate innovation from all areas of a business. Non-technical people who can clearly define the problem, desired outcome, constraints, and KPIs will be able to build tools using AI-enhanced development platforms without writing a single line of code. This will allow technical teams to focus on hard problems like architecture, reliability, and security, while non-technical builders turn ideas into working products much faster than before.

When I zoom out, all three of these predictions point in the same direction. AI usage becomes an expectation, job descriptions shift toward people who can work alongside automation, and non-technical builders gain leverage if they can turn clear ideas into working tools. These shifts are already in motion, and we are either learning how to work with them or pretending they are still optional. You do not control the timing, but you do control whether you arrive unprepared or with receipts. Start building those receipts now, one real use case at a time, so that when 2026 shows up you are not asking if AI will change your work, you are showing how you already changed with it.