This week at Hitachi Digital Services we published the “Skills are all you need” whitepaper. It starts from a blunt idea: enterprises are scaling AI the wrong way when they treat agents as the product, because that path usually creates more orchestration than ownership. Our argument is that the real unit of scale is the skill: bounded, governed and measurable. That lens made the rest of this week’s reading click into place. Google’s TurboQuant suggests the next advantage may come less from model hype and more from serving efficiency, memory and retrieval economics. A sharp essay on LLM psychology is a useful reminder that fluent systems make people lower their guard and over-credit what is often just polished output. Nicolas Bustamante’s take on vertical SaaS pushes the point further: if chat collapses learned interfaces and workflow lock-in, then some “moats” were really just expensive habits. And Joan Westenberg’s broadside against collaboration culture lands for the same reason: too many organisations confuse coordination with progress, just as too many AI stacks confuse agentic flexibility with real scalability.
In this issue
5 piecesSkills are all you need, is a recent whitepaper that I authored, arguing that enterprises are scaling AI the wrong way when they treat agents as the main delivery unit.
Our proposed alternative is the “skill”: a bounded, callable capability with a stable contract covering inputs, outputs, error handling, policy requirements, telemetry, cost and lifecycle. This separates open-ended reasoning from deterministic execution, reducing duplicated connectors, inconsistent governance and repeated integration work across multiple agent projects. We extend that argument into economics and security, as token spend is too narrow a cost metric for enterprise AI, because agent loops create volatile costs, whereas skills can declare p50, p95 and maximum cost envelopes per invocation. We detail the skill gateway, as a structural defence against prompt injection, lateral movement and supply-chain risk, then ties the model to MCP as an interoperability layer.
If you're still stuck at the agent unit and not the skills, then the consequence here is organisational, not technical. A lot of current enterprise agent work is building tomorrow’s integration estate by accident: the same systems connected three times, the same policy logic interpreted three ways, the same audit headache deferred until someone important asks awkward questions. There are plenty of agent demos still smuggling the hard bits inside the demo itself - we all been there! Governance disappears into glue code. Costs look acceptable because nobody’s forcing the system to declare its worst day, only its best screenshot. Bounded execution should be treated differently from open-ended reasoning, this is a grown-up architectural distinction, and more vendors should make it... Skills, in our framing, are basically productised enterprise capabilities. So, once you package execution behind contracts, manifests and gateways, you shift power away from whichever agent framework is fashionable this quarter and back toward the platform owner who controls access, policy and telemetry. There's a developer whitepaper too, on how to implement this idea.
Google Research’s Amir Zandieh and Vahab Mirrokni present TurboQuant as a quantisation method for large language models and vector search that aims to remove the usual memory overhead of compression.
The article says TurboQuant combines PolarQuant for the main compression step with QJL for a one-bit residual correction step, and claims this delivers large reductions in key-value cache size without accuracy loss. The write-up positions the work as both a practical systems improvement and a theoretically grounded algorithmic advance. The broader implication, as Google frames it, is cheaper and faster long-context inference plus more efficient high-dimensional vector retrieval.
Zero accuracy loss at 3 bits is the line that jumps off the page! That’s the kind of claim that makes infrastructure people lean forward, then immediately reach for the benchmark appendix and VC's fearing their investment... Google may well have something real here, as we've seen already implementation of this with Gemma4 on a MBP M4 running smoothly. The article does a good job explaining the core trick: spend most of the bits on a rotated, structured compression pass, then use a one-bit residual correction to clean up the damage. Elegant. Suspiciously elegant. There’s also a bigger systems point lurking underneath. Compression schemes that preserve quality without retraining don’t just save memory; they change product economics. Longer context windows become cheaper to serve. Retrieval-heavy architectures get more attractive. Search infrastructure starts to look more strategically important than another headline model release. That matters because the enterprises with the best serving stack, not merely the flashiest model, may end up with the most defensible advantage.
This essay argues that large language models are unusually hard to judge because people keep describing them with the language of minds, thought and intention.
The author says that framing distorts evaluation from the start, since LLMs are statistical systems rather than reasoning entities. Across sections on linguistic framing, pareidolia, chat interfaces and “Talking Dog Syndrome”, the piece claims that users routinely over-credit these systems simply because they resemble conversation and mimic human output. The article’s wider point is that the social wrapper around LLMs changes how people assess reliability. It argues that chat interfaces make users feel like collaborators, which lowers scepticism once a response becomes merely acceptable.
Great article and gets right into the psychological trap: on how quickly people slide from “this produced language” to “this understands language”, and that distinction matters more than a lot of product demos would like. Once users start iterating with a system, they begin rooting for it. That’s a real observation, and it explains plenty about why weak output gets waved through. What it leaves out is proof. There’s a lot of confidence here, sometimes more than argument. Saying LLMs are “pure mimicry” may be directionally consistent with the essay’s view, but the piece often prefers forceful phrasing over showing its working. “Talking rock” is memorable. It’s also a rhetorical device doing plenty of labour. Pratchett earns the reference. The point lands. The real complaint isn’t only that people misunderstand LLMs. It’s that the surrounding presentation is built to help them misunderstand. Still much more useful than another breathless piece about machine “reasoning”.
This X post argues that large language models are repricing the vertical software market by dismantling several of the moats that once justified high valuations and sticky customer bases.
Nicolas Bustamante draws on his experience at Doctrine and Fintool to claim that learned interfaces, encoded workflows, public-data access, talent scarcity and bundling are all being weakened or destroyed by LLMs. In his telling, natural-language interfaces, agentic workflows and markdown-defined skills collapse years of specialised product and engineering effort into a much cheaper, faster software stack.
Bustamante is really arguing that a chunk of vertical SaaS has been over-earning on interface friction, implementation pain and encoded workflow complexity, and that LLMs are stripping those premiums out faster than incumbents would like to admit. That’s a serious and interesting view... also an uncomfortable one if you’ve spent years mistaking customer entrenchment for customer love. He gets a lot right, must admit! The point about learned interfaces is especially strong because it names something buyers and builders both know but rarely say aloud: plenty of enterprise software is defended by habit, not delight. Same for the observation that domain expertise can now be turned into executable guidance far more directly than before. You can see why that changes the economics of who gets to build. Yet, agents don’t remove accountability, they relocate it... Strong diagnosis, slightly theatrical execution, and probably closer to true than most incumbents will say in public.
JA Westenberg argues that modern workplace collaboration is less a productivity engine than a system for diffusing responsibility.
The article ties that claim to examples ranging from S.L.A. Marshall’s disputed wartime observations to Ringelmann’s work on social loafing and Frederick Brooks’s software project thesis. Its central contention is that organisations now mistake coordination, visibility and participation for actual output, creating elaborate rituals of teamwork that obscure who is truly accountable for decisions and delivery. The piece extends that argument into a critique of the software and culture built around knowledge work. Slack, Jira, Notion, ClickUp and similar tools are presented as infrastructure for managing the appearance of collective effort rather than producing finished work. Westenberg does not reject communication outright, but argues that communication should support clear ownership by individuals or very small groups. The article’s broader implication is that collaboration-first cultures have become expensive accountability-avoidance machines.
Brooks is the tell here, right?! This article repeats a very old management pattern: institutions discover that adding process to fix coordination failure often creates a second, larger failure made of meetings, status theatre and procedural lint. Westenberg’s version is sharper, ruder and more fun to read than most - must admit! Plenty of teams do confuse visibility with progress and inclusion with ownership. You can feel that sentence in your bones if you’ve sat through a roadmap meeting that generated three follow-ups and no decision. Complex organisations do need coordination, they just don’t need coordination to become the product... There’s a meaningful difference between sharing context and building a bureaucracy of perpetual synchronisation. Small teams with clear authority usually do outperform diffuse groups on ambiguous work, and the orgs that remember that will move faster than the ones still polishing their collaboration stack... Just saying...
Last Week in AI
- AI overly affirms users asking for personal advice | news.stanford.edu
- State of Context Engineering in 2026 | newsletter.swirlai.com
- 9 Observations from Building with AI Agents | tomtunguz.com
- Developer’s Guide to Building ADK Agents with Skills | developers.googleblog.com
- Sandboxing AI agents, 100x faster | blog.cloudflare.com
- Gemma 4: Our most capable open models to date | blog.google