Divya van Mahajan

My Autumn AI Predictions

ai-daily-brief-podcast

Autumn AI Predictions: Study Document

Overview

This episode of the AI Daily Brief — a daily podcast and video covering AI news and analysis — presents the host’s forward-looking predictions for the AI industry heading into autumn 2025. The host (not named explicitly in the transcript) uses the “back to school” period as a natural inflection point to assess where the AI industry stands and forecast how key themes will develop over the coming months. The talk covers model developments, enterprise adoption, multimodal AI, cost dynamics, data sourcing, and industry consolidation.

Source video: Not available (transcript provided directly; no URL supplied)

Prerequisites

  • Familiarity with large language models (LLMs) and foundational AI companies (OpenAI, Anthropic, Google DeepMind, xAI, Meta)
  • Basic understanding of the AI product landscape: ChatGPT, Claude, Gemini, Grok
  • Awareness of enterprise AI adoption concepts: pilots, proofs of concept, agentic workflows
  • General knowledge of financial markets and venture capital dynamics
  • Some exposure to AI infrastructure concepts: inference costs, training data, model scaling

Main Points

1. AI Skepticism Is Simmering but Overstated

  • A widely circulated MIT study claiming 95% of AI pilots create no value drove a wave of AI skepticism, amplified by headline writers, market short-sellers, and media editors.
  • The study’s methodology was weak: approximately 50 interviews and readings of public financial statements, not direct value measurement.
  • The skepticism was amplified by summer market conditions — historically low liquidity, anxiety around rate-cutting cycles, and tariff policy uncertainty — making narratives more impactful than they would otherwise be.
  • Sam Altman’s comments about a possible AI bubble were widely over-reported and taken out of context.
  • Prediction: This skepticism narrative will fade, especially if markets receive an expected rate cut in September 2025. It will later feel less significant than it appeared when it broke.

2. Enterprise Focus Will Shift to Implementation Infrastructure

  • The MIT study, combined with GPT-5’s launch, is pushing enterprises to think less about model performance in isolation and more about the hard work of implementation.
  • The Harvard Business Review published a piece on escaping the “AI experimentation trap,” advising companies to focus on high-value use cases rather than spreading too thin — a sign of where the enterprise conversation is heading.
  • Prediction: The fall will see increased focus on context orchestration/context engineering — ensuring AI agents have access to relevant organisational data and background to function effectively.
  • MCP (Model Context Protocol) has been a major 2025 theme; A2A (Agent-to-Agent) workflows are predicted to dominate 2026 headlines, per Ravine (CEO of Cradle).
  • Single agents are insufficient for enterprise complexity; specialised agents communicating via A2A standards will become the next infrastructure priority.

3. Counter-Signals to Skepticism: Market and Fundraising Momentum

  • Evercore bucked the bubble narrative with a note predicting a 20% US stock rally by 2026, driven in part by AI.
  • Recent financial results: CoreWeave tripled revenue; Microsoft Azure grew 39%; NVIDIA grew 56%.
  • Anthropic closed a funding round at a $183 billion valuation, raising $13 billion — three to four times its originally anticipated raise — after 5x-ing revenue from $1B annualised (January) to $5B annualised (August 2025).
  • Vibe-coding platform Lovable is receiving acquisition offers at a $4 billion valuation, up from a $1.8 billion raise just prior.
  • Prediction: Private investor enthusiasm for AI deals will continue to grow unabated.

4. Agentic Coding Is the Defining AI Use Case of 2025

  • “Vibe coding” was not a term at the start of 2025; by mid-year it had become a dominant force.
  • Lovable jumped to #23 in Andreessen Horowitz’s top 100 Gen AI consumer apps, with traffic growth to both platform sites and user-created domains — indicating real production use.
  • Prediction: The term “vibe coding” will be replaced by “agentic coding” as a more accurate descriptor of the full range of use cases, including professional developers.
  • Agentic coding will move out of the prototype phase inside enterprises and into normalised production workflows.
  • Supporting tooling will grow, e.g., Lindy’s code review agent for agentic-coded applications.
  • Prediction: By end-of-year lists, agentic coding will be recognised as the most important AI force of 2025.

5. Multimodal AI Has Been Quietly Exploding

  • While attention focused on LLM chatbot benchmarks, multimodal capabilities have advanced dramatically.
  • NanoBanana (technically Gemini 2.5 Flash Image): excels at making precise edits to existing photos, unlocking many commercially viable use cases not previously possible.
  • VO3: first video-generation model to reach production for national TV advertising.
  • Genie 3 (previewed August 2025): major advances in 3D interactive world generation from text prompts, with extended memory.
  • Tools like Higgs Field are being combined with image/video models to create polished AI video content.
  • Prediction: As summer debates fade, users and companies will recognise an expanded creative canvas they have barely explored, and multimodal capabilities will move more into production.

6. AI Costs Are Simultaneously Falling and Rising (Jevons Paradox)

  • Inference costs are falling at a rate exceeding Moore’s Law comparisons, yet total AI spend continues to rise.
  • Explanation: Aaron Levy (Box) described this as Jevons Paradox — as tokens get cheaper and more capable, organisations use far more of them for increasingly complex tasks.
  • Example: Getting a 99% correct legal contract answer vs. 90% may justify a 10x–100x increase in token usage.
  • Over time, specific task categories will plateau and see real cost reductions; but new higher-complexity uses will continuously absorb freed-up capacity.
  • Agentic coding exemplifies this: multiple coding agents running in parallel consume vastly more tokens than single-turn interactions.
  • Prediction: Total AI spend will continue to rise in the near term even as per-unit costs fall, before eventually specific workload categories become materially cheaper.

7. New Data Sources Become a Strategic Priority

  • Pre-training scaling as a methodology may be approaching diminishing returns; models have now been trained on essentially all publicly available human text.
  • Prediction: AI labs will increasingly seek novel, proprietary data sources.
  • Anthropic updated its terms of service to make consumer interaction data opt-out rather than opt-in for training.
  • OpenAI and xAI have discussed licensing or purchasing Cursor’s coding interaction data — millions of software engineers’ real-world coding behaviour.
  • Prediction: Labs may pursue deals with enterprise content management providers for access to proprietary enterprise data.

8. Major New Model Releases Are Unlikely in 2025

  • Sam Altman has referenced GPT-6 as coming faster than the GPT-4 to GPT-5 gap, but has walked back any suggestion of an imminent release; memory is flagged as a key feature emphasis.
  • Elon Musk has indicated intent to release Grok 5 by year-end.
  • Gemini 3 rumours are internet-driven, not Google-encouraged.
  • Prediction: No major flagship model releases in calendar year 2025. Labs will wait until they have an incontrovertibly clear advancement, having observed mixed initial reactions to GPT-5.
  • Ranked order of likelihood for 2025 release: Grok 5 > Gemini 3 > GPT-6.
  • Expect instead: multimodal updates and open-weight model advances, particularly from Chinese companies.

9. Foundation Model Labs Are Shifting Toward Applications

  • OpenAI hired Fiji Simo (former Instacart CEO) as CEO of Applications, focused on practical ChatGPT use cases.
  • OpenAI acquired a product testing startup; OpenAI CEO Kevin Weil announced “OpenAI for Science” — an AI platform for accelerating scientific discovery.
  • Anthropic released Claude for Financial Services — a custom-branded product for a specific vertical.
  • Prediction: Foundation model companies will increasingly build or acquire vertical-specific applications.
  • Key open question: which verticals will foundation model companies compete in directly, and what space remains for third-party vertical AI apps?
  • The host is cautiously optimistic that third-party apps can survive via contextual industry knowledge, proprietary data, and UX specialisation — but competitive dynamics will shift.

10. Consolidation: At Least One Major M&A Deal Expected

  • The AI landscape has not seen a major structural shift since Inflection/Microsoft and Character AI/Google (both over a year ago).
  • Apple has been internally pushed toward acquisitions (Mistral and Perplexity cited as targets), but Tim Cook has reportedly blocked these.
  • Meta (Zuckerberg) is spending heavily — including a semi-acquisition of Scale AI to secure Alexander Wang — and is a credible acquirer.
  • Goldman Sachs is predicting record M&A broadly in 2026.
  • Prediction: At least one major unexpected M&A deal will occur this autumn, but it will not involve Apple. Apple is unlikely to embrace the cultural and strategic shift required.

Key Concepts

  • MCP (Model Context Protocol): A standard for how AI agents access external tools, data, and context within enterprise systems; a major theme of 2025.
  • A2A (Agent-to-Agent) workflows: Communication standards enabling specialised AI agents to collaborate on complex tasks, predicted to dominate 2026.
  • Context orchestration / context engineering: The discipline of ensuring AI agents have access to the relevant organisational data and background needed to perform effectively; framed as the next major enterprise AI theme.
  • Agentic coding: The use of AI agents — potentially running in parallel — to write, edit, review, and deploy code, often autonomously; evolution of “vibe coding.”
  • Vibe coding: Informal term for non-engineers using AI tools to build software applications; expected to be superseded by the broader term “agentic coding.”
  • Jevons Paradox: An economics principle stating that as a resource becomes more efficient and cheaper, total consumption of that resource tends to increase rather than decrease.
  • NanoBanana (Gemini 2.5 Flash Image): Google’s image editing model notable for precise, targeted modifications to existing photos rather than raw generative power.
  • VO3: Google’s video generation model; the first to reach production use in nationally aired television advertising.
  • Genie 3: Google DeepMind’s world model capable of generating 3D interactive environments from text prompts with extended memory.
  • Pre-training scaling: The methodology of improving LLM capability by training on ever-larger datasets using more compute; subject to debate about whether diminishing returns have been reached.
  • Inference cost: The computational cost of running a trained AI model to produce outputs; has been falling dramatically faster than predicted.

Summary

The host argues that the wave of AI skepticism in summer 2025 — driven by the MIT pilot study, market conditions, and over-reported commentary from Sam Altman — is contextually inflated and will fade as markets stabilise and strong financial results from AI-exposed companies reassert themselves. Beneath the noise, the host sees several durable trends accelerating into autumn: enterprise AI focus is shifting from flashy pilots to the difficult infrastructure work of context engineering and agent interoperability (MCP, A2A); agentic coding has emerged as the defining AI use case of 2025 and will only deepen; multimodal capabilities have been quietly advancing at an exponential pace and are entering real production; and total AI spend will continue rising even as unit costs fall, due to Jevons Paradox dynamics. Looking ahead, the host does not expect flagship new models (GPT-6, Grok 5, Gemini 3) to ship in 2025, but anticipates foundation model labs pivoting more visibly toward vertical applications, at least one significant M&A deal surprising the market, and a growing scramble by labs to acquire novel proprietary data as public training corpora are exhausted.