The BlueNexus team are constantly researching emerging trends within the AI sector. Earlier this week we came across an extremely interesting article which proposed the notion of focusing strongly on context within LLM training methods. We find this particularly interesting as it strongly aligns with our product offering and wider vision of how AI not only should be developed, how it must be developed.

What if the secret to building smarter AI agents wasn’t better models, but rather better memory & context? This is the core idea behind Yichao Ji’s recent writeup, which details lessons from developing Manus, a production-grade AI system that ditched traditional model training in favour of something far more agile - "context engineering".

From Training to Thinking

Rather than teaching an LLM what to think through intensive fine-tuning, Manus has been focusing on designing how it thinks, via structured, persistent, runtime context.

Key tactics include:

• KV-cache optimization to reduce latency and cost
• External memory layers that store files and tasks without bloating prompts
• Contextual “recitation”, for example agents reminding themselves of their to-do list
• Error preservation as a learning loop
• Tool masking over tool removal, to retain compatibility and stability

This approach points to a deeper shift in the LLM training debate, shifting from “prompt engineering” to context architecture, and it’s changing how intelligent systems are being built.

Diving Deeper

Ji’s article observes that developers still default to the “If the model isn’t good enough, retrain it" approach. But Manus demonstrates that this isn't scalable. It’s expensive, brittle, & hard to maintain across use cases. Instead, they show that by designing the right context window with memory, goals, state, & constraints, developer you can achieve robust agentic behavior 𝐟𝐫𝐨𝐦 𝐞𝐱𝐢𝐬𝐭𝐢𝐧𝐠 𝐋𝐋𝐌𝐬.

We don't necessarily see this as a "work around" but rather the new standard emerging, which is fantastic within the R&D lens of LLM training.

Obligatory mention that we carry some level of bias here, as this new standard plays straight into our wheelhouse.

Alas, BlueNexus Agrees

We wont sit here and "Shill" this approach from the roof tops, it’s fair to say this emerging standard aligns strongly with what we have been building.

The future of AI isn’t just about inference, speed or model accuracy, in our opinion it’s about relevance, continuity, portability and coordination.

By this we mean:

• Knowing what data should (and shouldn’t) be in scope within any given prompt or automation
• Remembering past actions across sessions & various tools / 3rd party applications
• Structuring memory & state for reasoning, not just retrieval

As always, were interested in what other AI builders think?

• Are we overvaluing model complexity & undervaluing memory infrastructure?
• What makes context trustworthy, especially across tools, users, & time?
• Could context-based architectures unlock broader access to AI, without the cost of custom training?
• Is “context as code” the new OS for agents?

We would love to get a collective thoughts across the spectrum from anyone participating in this space. Feel free to add your colour to the conversation & start a dialogue with likeminded people in the comments below.

Quick one today, as the team focuses on final checks and balances for an upcoming deployment. In this weeks article, we examine old business models coming to an end as new tech muscles its way through the "walled gargens" of big tech, to create a more open and interoperable future.

For years, Big Tech has profited by building "walled gardens", that is - closed platforms that tightly control apps, data, and user experience. These ecosystems, like Apple’s App Store or Meta’s social networks, maximize revenue and lock-in but restrict user freedom and innovation. Today, the tide is turning. Consumers, regulators, and even some companies are now demanding interoperability: seamless connection between apps, platforms, and devices. The result? A major shift in how digital infrastructure is being built.

The Signs of Change

1. ‍Cross-Platform Messaging: Apple will adopt RCS, improving messaging between iPhones and Androids. Under EU pressure, Meta is working on interoperable WhatsApp and Messenger platforms. The days of siloed chat apps are numbered.
2. ‍Smart Home Unification: Matter, an open smart home protocol supported by Apple, Google, Amazon, and others, now enables devices to work together regardless of brand.
3. ‍App Store Alternatives: In response to Europe’s Digital Markets Act (DMA), Apple is allowing third-party app stores and alternative payment systems on iOS—a sea change in how apps are distributed.
4. ‍Data Portability: New rules in the EU require platforms to let users move their data between services. Companies like Meta and Google now offer improved tools for exporting content and contacts.
5. ‍Open Social Networks: Meta’s Threads plans to integrate with ActivityPub, allowing users to interact across decentralized platforms. Tumblr and Flipboard are doing the same.

Consumer Preference

Interoperability matters to users because it offers:

• Freedom to switch platforms without losing data.
• Convenience of unified experiences across devices.
• Transparency and trust through open standards.

Surveys show over 80% of consumers prefer tech that plays well with others. In areas like smart homes and messaging, users are tired of being forced into one brand's system.

Why Companies Are Opening Up

• Regulation: Laws like the DMA in Europe are forcing change.
• User Expectations: Consumers demand flexibility and integration.
• Strategic Advantage: Interoperability can reduce churn and expand markets.

Even historically closed companies are making adjustments. Apple, for example, is embracing RCS and opening up iOS in Europe.

The Challenges Ahead

Interoperability isn’t without hurdles:

• Security and privacy risks increase when platforms open up.
• Loss of revenue from closed ecosystems (e.g., Apple’s App Store fees).
• Fragmentation if standards aren’t widely adopted or well managed.

The Bigger Picture

This is about more than messaging or app stores. It’s about redefining how digital systems are structured: from isolated silos to connected ecosystems. Users will gain freedom, developers will gain reach, and companies that embrace openness early may gain a competitive edge.

Final Thought

The push for interoperability is more than a trend it’s a structural shift. Platforms that prioritize openness, data portability, and integration are better positioned for the future. For users and builders alike, the walls are coming down—and the digital world is becoming a more open place to build.

Since the inception of the first mainstream retail facing AI (GPT), the dominant business model for AI assistants has been paywalls (Pro tiers) and usage-based APIs. But as inference costs fall, LLM models converge in their abilities and assistants eat more of the consumer attention stack, signs point to a familiar destination: ads.

A Race to the Bottom?

Three forces are converging where LLM’s are concerned.

1. Rapid price compression - Analyses from a16z and others show LLM inference costs collapsing at extraordinary rates (10× per year for equivalent performance in some estimates), which pressures providers to cut prices to stay competitive and expand usage footprints. Over time, cheaper inference makes ad-supported models more viable at massive scale.
2. Platforms are already testing ads in AI UX. Perplexity began experimenting with ads (including “sponsored questions”), laying out formats that blend with conversational answers. Google now shows Search/Shopping/App ads above or below AI Overviews, and leadership has telegraphed “very good ideas” for Gemini-native ads. That’s a direct bridge from keyword ads to AI answers. Snap and others are rolling out AI-driven ad formats (sponsored Lenses, inbox ads), normalizing AI-mediated, personalized placements.
3. The search precedent. Ad-free, subscription search (Neeva) closed its consumer product, an instructive data point about the difficulty of funding broad information services purely with subscriptions.

Put together: the economics and UX rails for advertising inside assistants are falling into place.

But it’s not that simple: 3 strategic counter-currents

A. API revenue isn’t going away. Enterprise APIs remain sticky, and top-tier reasoning models still carry non-trivial costs (driving usage-based pricing and value-based packaging). Even bullish observers note advanced tasks incur higher costs that won’t commoditize as quickly.

B. Regulation & trust are tightening. The FTC is actively targeting deceptive AI advertising and claims, and California’s CPRA expands opt-outs and limits around sensitive data—guardrails that complicate hyper-targeted ads based on AI-enriched profiles.

C. Cookies aren’t (fully) dead, yet. Google’s third-party-cookie phase-out has been delayed and reshaped multiple times, signaling a messy transition from old targeting rails to new ones. That uncertainty slows the clean hand-off to purely AI-native ad targeting.

The likely outcome, a “tri-monetization” model

Expect leading AI platforms to run three parallel models:

1. Consumer Free + Ads. Assistants inject sponsored answers, product placements, or commerce links—especially in high-intent categories (travel, shopping, local). This aligns with how Google is already positioning ads around AI Overviews and how Perplexity has tested formats. There are some nuances here which will all come down to delivery and execution.
2. Premium Subscriptions. Ad-light or ad-free tiers with priority compute, longer context windows, and premium tools (collaboration, analytics). Even if ads expand, a sizable cohort will pay to reduce ad load and raise limits, similar to the Spotify playbook.
3. Enterprise SaaS + Usage-Based APIs. The durable, high-margin layer: SLAs, governance, connectors, private deployment options, and compliance guarantees. This remains where buyers pay for certainty (and where ad models don’t fit).

The interesting notion about this prospective shift in revenue models is how the wider retail market will react.

Consumers have become so accustomed to the “Data Stockholm model” - the long-standing trade of free software for personal data — that it has evolved into a kind of digital cultural norm. For decades, people have accepted the idea that access to “free” platforms comes at the hidden cost of surveillance, profiling, and monetization of their digital selves.

That uneasy equilibrium mostly held when the algorithms behind those systems were static and predictable. But as AI becomes the interface for nearly every digital interaction, the equation changes. The idea of handing over your personal data not to a dumb algorithm, but to a self-learning system capable of generating, inferring, and acting on that data introduces an new layer of discomfort.

Public trust in big tech is already fraying. Recent surveys show a majority of users are uneasy about companies using personal data to train generative models. This raises a crucial question:

Are consumers ready to pay for AI services in exchange for real privacy and data autonomy?

Or will they continue to tolerate the invisible bargain - accepting “free” AI assistants that quietly harvest behavioural data to fuel model training and hyper-personalized advertising?

While many retail users may not fully grasp the nuanced implications of AI-driven data use, the notion of data sovereignty - owning and controlling your own digital footprint, is beginning to resonate. It may well become the catalyst for a cultural shift: away from “free for data” toward paid trust.

If that shift happens, it won’t just redefine how AI is monetized; it will redefine how digital trust itself is valued.

Hyper-personalized ads: the promise and the peril

Should retail choose to continue with the status quo, lets examine how that may look. Firstly, why would such large players such as open AI & Anthropic even consider adding advertising to the mix, thats an instant turn off, right ? The issue isn’t necessarily wether this is an intentional choice, but rather a financially strategic play. For example while OpenAI boast am impressive MAU of 800 Million users (just under 1/4 of the global populace), only 5% of those users are paying users. When we couple this figure in with the fact that OpenAI carried a 5 billion dollar loss in 2024 (forecasted to be as high as 14 billion by 2026), it is clear that there is going to be an uphill battle to condition consumer behaviour away from the “free for data” mindset & towards a more traditional monetary exchange model.

This notion is amplified, when we question what LLM’s will look like in the next decade. Some argue that this is a virtual “race to the bottom”, where by LLM models will eventually offer little distinction from one and other, thus, the battler for market share wont come down to product, but price. As this “digital Mexican stand off” takes effect, it will all come down to who blinks first. When one constructs these three factors into a logical argument for business strategy, its not too far fetched to conclude that most LLM’s will end up generating a vast majority of their revenue, from advertising, carefully curates and served up courtsey of the data that users are feeding their preferred LLM model.

If this becomes a norm, is it really all that bad? Lets quickly examine the pros and cons.

Pros. AI’s ability to model real-time context could make ads more useful: for example, an assistant that (with consent) knows your itinerary, food allergies, and budget to surface the right restaurant with instant booking. WPP’s recent $400m partnership with Google is a great example of how agencies are betting on AI-scaled personalisation and creative generation.

Cons. Hyper-personalization relies on first-party data and sensitive signals. While there are regulatory and legislative limits in place for the use of sensitive personal information, these protections are geographically skewed and have alot of catching up to do. Until such guardrails are put in place, the protection of how your data is used, primarily comes down to individual product “terms of use” - When was the last time you read one of those?

Its the Users Choice, but Chose Carefully

If hyper-personalized ads are inevitable in some AI contexts, they must be consented, sovereign, and provable. Architectures that keep user data in controlled environments, attach revocable consent to every data field, and log every use give enterprises a way to experiment with monetization without corroding trust. That’s where privacy-preserving data vaults, confidential compute, and auditable policy enforcement become not just “nice to have,” but the business enabler.

Question for readers: If assistants start funding themselves with ads, what standards (disclosure, consent, data boundaries) should be mandatory before you’d allow them to personalize offers for your customers or you as a retail user?