AI In 2027: What I Expect, What I’m Unsure About, And 10 Predictions

🔍 Read the full analysis: AI In 2027: What I Expect, What I’m Unsure About, And 10 Predictions on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

Thorsten Meyer published 10 predictions for AI in 2027 on October 15, assigning each a confidence level and describing signals he plans to use to judge whether it comes true. The essay forecasts a shift from one-question-at-a-time chat toward longer-running agents, AI-generated code and persistent work memory, while identifying reliability, verification and electricity supply as constraints; the predictions remain the author’s outlook, not confirmed developments, in a field also explored in Kevin Kelly’s AI predictions.

Meyer says his predictions have an average confidence rating of 72.5%. He frames that number as a way to make the forecast testable: by October 2027, he expects roughly seven predictions to be borne out if his confidence is well calibrated. That is his stated scoring framework, not an independently assessed probability or a guarantee that seven outcomes will occur.

Among the highest-rated calls, Meyer gives 85% confidence that AI assistants will retain information about users’ projects, decisions and open questions across conversations. He assigns 80% to a growing emphasis on evidence—such as tests, sources and traces—to check AI-generated work, and to power and computing capacity becoming limits on what AI services can deliver. He also predicts that routine assistant tasks will increasingly run on personal devices rather than relying entirely on cloud computing.

Other forecasts include agents handling work over several days, more production code being generated by AI, and AI systems assisting scientific research. Meyer assigns 70% confidence to multi-day delegation, 75% to most new code no longer being typed by people, and 65% to AI becoming a lab partner. The supplied excerpt begins describing the research prediction but cuts off before completing it and does not include the remaining four predictions, so their substance and ratings cannot be reported here.

From Chat Prompts to Delegated Work

The forecasts point to a practical change in how people may use AI: less time composing individual prompts and more time setting goals, checking intermediate work and deciding whether results are acceptable. If longer-running agents become dependable, people could hand over sequences of tasks rather than supervise every step. Meyer’s own caveat is that agents can drift from the original goal over time, making visibility into their work important.

The emphasis on verification also has implications for organizations adopting AI. A plausible answer or a program that compiles may not be enough to establish that a result is correct. Meyer argues that tests, evidence and independent review could become part of routine workflows. This is a prediction about practices he expects to spread, not evidence that organizations have already adopted them widely.

Persistent memory and local computing raise separate questions for users and buyers. An assistant that remembers work could reduce repeated explanations, but that memory may contain sensitive details. At the same time, running more tasks on devices could reduce dependence on data centers for routine work, though the essay does not quantify possible performance, cost or privacy effects.

How Meyer Plans to Score the Forecast

Meyer describes the piece as a forecast, not a prophecy. He says each prediction has a confidence rating and an infographic containing signals he will use to assess it. He also sets October 2027 as the point by which he expects to review the calls. The supplied source says a short video accompanies the essay, but the article’s factual basis here is the written material provided.

The predictions draw on arguments Meyer says he has made in earlier writing about AI demand, energy constraints and coding work. In this essay, he links advances in AI software to physical limits such as electricity, chips and cooling, and to the need for clearly specified tasks and reliable tests. These are the author’s explanations for why he expects particular changes; the source does not provide independent measurements or third-party forecasts to verify them.

The source material supplied for this article contains only the first six prediction headings and partial detail on the sixth. Although the essay is described as containing ten predictions, the remaining four are not available in the material, and their claims should not be inferred from the opening list.

“A forecast, not a prophecy.”

— Thorsten Meyer, in the October 15, 2026, essay

Reliability, Power and Missing Predictions

None of the 2027 outcomes in the essay is confirmed. Meyer says agent reliability remains unresolved, particularly the risk that a system working over several days may drift from its assignment. The source does not provide benchmark results showing how often this happens, nor does it establish how quickly developers may address it.

It is also unclear how fast electricity, chips and cooling will constrain AI services, or how much routine work will move onto phones and laptops. Meyer offers these as forecast drivers but supplies no capacity figures or quantified timeline in the provided material. His proposed user control over AI memory—inspection, editing, export and choice of storage location—is likewise an expectation, not a description of a standard already in place.

Finally, the source excerpt does not include the full text of predictions seven through ten and stops partway through the research forecast. Their content, confidence ratings and scoring criteria are not available in the material provided. The 72.5% average applies to all ten ratings as Meyer reports them, but the excerpt does not allow readers to review all ten calls.

The October 2027 Scorecard

Meyer says he will judge the predictions by the signals listed beneath each item and revisit the results by October 2027. That review is the next stated milestone for the forecast. The supplied source does not specify an exact publication date for the scorecard or explain how conflicting evidence will be handled.

For readers tracking the claims, the most direct tests will be whether agents are routinely trusted with multi-day work, whether AI-generated code becomes the main source of new production code, and whether assistants retain useful work context across sessions. Evidence of adoption will matter alongside product demonstrations: Meyer’s own criteria distinguish systems that work in practice from features that remain demos.

Readers will also need to watch for changes in verification practices and computing infrastructure. Whether AI services increasingly show tests and sources, and whether power constraints affect availability or price, remain open questions. The complete essay would be needed to assess the other four predictions and their individual evaluation signals.

Key Questions

What did Thorsten Meyer publish?

He published 10 predictions for AI in 2027, each with a confidence rating and, he says, signals for judging it. The source material provided here contains details for only the first six, with the sixth cut off partway through.

Are the predictions confirmed developments?

No. They are Meyer’s forecasts about what may happen by October 2027. The article describes expectations and uncertainties, not outcomes that have already been verified.

Which prediction did Meyer rate most confidently?

In the supplied excerpt, his highest rating is 85% for AI assistants retaining information about users’ work between conversations. Meyer also says users will have to decide where that memory is stored and who controls it.

What does Meyer say could limit AI progress?

He points to electricity, chips and cooling as potential limits on computing capacity. He also identifies agent reliability and the ability to verify AI-generated work as practical challenges.

When will the predictions be evaluated?

Meyer frames the forecast around an assessment by October 2027 and says each prediction has signals he will use to score it. The provided source does not give a more precise date or a method for resolving ambiguous results.

Source: ThorstenMeyerAI.com