Context home windows could turn into a bottleneck once more.
Context home windows have gotten enormous during the last six months — we’re already seeing context home windows of 1 million, 2 million, even 10 million tokens — however our aspirations have grown too. We need to feed fashions even bigger troves of paperwork, codebases, movies. Whereas brokers are bettering at managing information bigger than their context window by partitioning, sampling, and summarization, we’re additionally operating into the boundaries of how properly fashions can deal with ultra-long contexts with out shedding accuracy or growing prices. With exponentially extra content material being generated by GenAI itself, it’s very doable we’ll discover ourselves bottlenecked once more, even with these expanded capacities.
Knowledge science and conventional ML will turn into commonplace agent instruments.
That is one factor I’m very enthusiastic about. It is going to turn into commonplace for AI frameworks to ship brokers that aren’t simply answering queries or producing textual content, however whipping up a linear regression pipeline, operating a clustering evaluation, or producing exploratory plots. Similar to climate or foreign money conversion instruments are commonplace, we’ll begin seeing a “information evaluation” device included by default. If you would like a superb instance of an agent that may do that, see Google’s samples of use for its new ADK framework.
RAG will fade, then come roaring again.
As context home windows bought enormous, it’s appeared like RAG (retrieval-augmented technology) was wanted in fewer circumstances. This can be a paradigm shift that just some have seen. However as a direct consequence of #1, RAG will begin to swing again to being indispensable for the preponderance of GenAI enterprise use circumstances once more. This can not less than have begun to occur by Might 2026. The important thing distinction is that this time round, we’ll have way more mature managed RAG options. Individuals will more and more notice they need to be utilizing managed RAG 90% of the time they’re utilizing RAG, reasonably than constructing all of it themselves.
AI will energy main scientific breakthroughs.
We’re already seeing AI uncover new antibiotics, optimize physics experiments, and design new molecules. Very thrilling!
Brokers will combine extra deeply with enterprise information.
We’re already seeing Google Workspace and Microsoft 365 combine AI throughout Gmail, Drive, Docs, Excel, and Groups — letting AI brokers question your organization’s belongings seamlessly. This development will proceed, with AI brokers changing into much more conscious of inner paperwork, emails, calendars, and information bases, making them higher grounded within the group’s precise information.
Slop will proceed.
Slop (low-quality synthetic intelligence-generated content material) isn’t going away anytime quickly. Platforms for neighborhood content material, e mail purchasers, and another kind of content material feed will rise and fall primarily based on how good they’re at not letting slop discover its manner into your feed and inbox.
Coding kinds and tradition will change as AI-first coding stacks and vibe-coding stacks turn into widespread.
The position of the developer will change. Instruments like Cursor are already actually good AI-first IDEs. Whereas they’re not excellent but, they’re already a transformative improve, and by 2026 they’ll be a seamless a part of coding for 10 occasions as many builders. Consequently, how we code will change in stunning methods. Feedback might be extra verbose. Code kinds will turn into extra uniform. The DRY precept will turn into much less well-liked. Information will turn into shorter. However there might be a bigger variety of information.
Software program engineers might be given extra time to upskill.
The millennial technology is the biggest one because the child boomers, and we’re getting older. As we become old and extra senior, we’re given extra freedom and autonomy to study. Moreover, as instruments and ecosystems evolve quickly, engineers might be extra productive (releasing up time), and issues are altering very quick — making it a necessity.
High-quality-tuned and quantized edge fashions will assist clear up many issues.
High-quality-tuning utilizing PEFT, LoRA, and prefix tuning will allow IT to co-locate a small mannequin in your firm’s servers, lowering latency. We are going to discover jobs for brokers that do one factor reliably. And we’ll make it so they’re very, very small.
One other chip scarcity will certainly come.
Even when tariffs exempt most {hardware}, the longer AI advances, the extra methods we’ll discover to throw further brokers at issues. With this shift of the demand curve, the availability curve will meet the demand curve at the next worth level.
We’ll see the largest AI-driven inventory market shock but.
It would right inside an hour, 24 hours, or linger longer, however one thing’s going to occur. It’s been too quiet on that entrance for the way a lot AI is already built-in into buying and selling.
The controversy over K2–18b will sharpen.
Going outdoors my lane for this final one! The dialogue over whether or not dimethyl sulfide in K2–18b’s ambiance is a biosignature will proceed, with new information clarifying the sign. A 12 months from now, we’re prone to have way more certainty: both stronger affirmation of this doable signal of alien life, or a refutation pointing to an abiotic rationalization.
I provide advising companies on all the above (besides exoplanets) — be happy to e mail me should you’re exploring this.
