The General Intelligence Monolith Is Getting Refactored

The industry’s preferred fairy tale is that intelligence arrives as a single gleaming object: first a chatbot, then an agent, then AGI, presumably with a tasteful launch video and an enterprise pricing tier. This week’s more useful story is almost the opposite. The economically interesting AI stack is being decomposed: compress the broad model, delegate routine choices to a specialised one, and bolt controls around everything that can embarrass Legal.

PrismML’s Bonsai 2 27B is a claim that a 27B-class multimodal model can shed most of its memory burden without shedding much reported benchmark performance. The company says its ternary-weight model occupies 5.9GB, uses 1.76 effective bits per weight, and posts an 83.9 aggregate benchmark score against 85.4 for the full-precision Qwen3.8 27B base. Those are vendor-reported numbers, not a neutral coronation, but the direction matters: model quality is increasingly a systems problem, not merely a parameter-count contest.

Meanwhile, TypeSafe AI’s Jev declines the invitation to write fluent paragraphs at all. It is intended to take workflow state and return a bounded decision plus probability: invoke this tool, approve that request, route the task elsewhere. That is less cinematic than a synthetic colleague holding forth in Slack. It is also much closer to what a great many production systems need. The universal intelligence layer may survive, but mainly as the expensive generalist called in when the boring work has already been made explicit.

Bonsai’s reported specification is a useful rebuke to the assumption that local capability requires either a rack of accelerators or ritual sacrifices to a cloud bill. PrismML says the model supports a 262K-token context window, accepts text and images, runs through CUDA on Nvidia hardware and through MLX on Apple devices, and is Apache 2.0 licensed. It reports throughput up to 143 tokens per second on an RTX 5090 and 46.8 on an M5 Max. If those figures transfer to real workloads, a serious model becomes deployable in places where “send the document to an external API” is not an architecture but a compliance incident waiting for a ticket number.

The phrase doing the most work here is near-lossless—indeed, PrismML also uses “lossless” in its own framing. Compression always has a bill; the question is where it appears. Aggregate benchmarks can conceal regressions in particular languages, tool-use edge cases, long-context retrieval, visual reasoning, or the pathological little inputs customers seem to produce solely to test whether your monitoring exists. The supplied announcement does not include the benchmark methodology, generation settings, batch sizes, or the basis of its aggregate retention calculation. A benchmark average is not a service-level agreement wearing glasses.

Still, the technical implication is broader than one release. Efficient weights change the deployment envelope, and deployment envelope changes governance options. A compact model can sit nearer private data, run in a constrained environment, or serve as the cheap first pass before a larger model handles ambiguity. That is a more credible hybrid architecture than declaring every spreadsheet and support queue to be an AGI research project. The frontier is not only moving outward; it is being folded, quantised, and pushed toward the edge.

Jev represents the other half of the refactor: stop asking a generative model to impersonate deterministic software, then acting surprised when it behaves like an improviser. TypeSafe says its model delivers bounded decisions in 70 to 500 milliseconds and prices input at $0.042 per million tokens. The latency and cost claims are the company’s own and lack published workload details, accuracy results, or production-customer evidence. But the design instinct is sound: a routing decision need not arrive in five eloquent paragraphs with an optional haiku.

The catch is that bounded does not mean accountable. As InfoWorld’s report notes, users must define the question, possible outputs, thresholds, and escalation paths up front—and validate accuracy and calibration as their data shifts. In other words, someone still has to understand the business process. AI has not eliminated workflow design; it has merely made the absence of workflow design executable at lower latency.

That is why the enterprise governance market is proliferating into catalogs, gateways, runtime filters, data-flow monitors, model inventories, governance graphs, and automated red teams. The roundup lists 16 vendors, but it is a market map rather than an independent bake-off. Its real message is less flattering: every layer of agency creates another surface to inventory, test, constrain, explain, and occasionally switch off. Buying a control plane does not confer control any more than buying observability makes the distributed system less haunted. It does, however, give the haunting somewhere to emit metrics.

There is an awkward material dependency beneath all this elegant systems design: models learn from work made by people and institutions that still need revenue. Newly unredacted allegations in The New York Times copyright case describe publishers’ material being scraped and used in AI training, alleged paywall circumvention, removal of copyright notices, and systems whose answers may substitute for clicks. TechCrunch’s account stresses that much of the material comes from a legal brief, underlying exhibits remain sealed, and these are allegations rather than court findings.

The filings reportedly include a Microsoft document with the unusually candid diagnosis that LLM products risk undermining the economic foundations of their own content suppliers. This is not merely a copyright dispute conducted through expensive PDFs. It is an incentive problem in its purest form: extract value from the source, make the interface increasingly substitutive, then discover that the source cannot afford to keep producing the thing being extracted. Silicon Valley has reinvented the depleted fishery, except the cod have lawyers and bylines.

The legal answer remains unsettled; the report says courts have often been receptive to AI firms’ fair-use arguments. But the engineering answer cannot be postponed indefinitely behind legal ambiguity. A model that is cheaper to run, more capable at answering directly, and embedded deeper into workflows intensifies the commercial pressure on upstream information producers. Efficient inference is good. Efficiently erasing the incentives for reliable source material is a fairly eccentric long-term data strategy.

Against this backdrop, claims that AGI has arrived or is imminent look less like a technical milestone than a category-management exercise. Scientific American’s examination of OpenAI’s post-Astra positioning finds no shared definition. Some researchers treat broad linguistic and reasoning capacity as general intelligence; others point to current systems’ inefficiency, lack of durable learning from ongoing experience, and questionable creativity. One can call that a philosophical disagreement. One can also call it the absence of an acceptance test.

For CTOs and ML teams, the operational conclusion is satisfyingly unromantic. Do not architect around an ontological announcement. Architect around workloads, failure modes, latency budgets, data boundaries, evaluation harnesses, and the human authority required when the system is uncertain. A compressed general-purpose model may be the right private analyst. A specialised decision model may be the right router. A traditional rules engine may still be the right approver, which is not a defeat for progress but a victory for not using a chainsaw to butter toast.

The serious AI organisation will therefore look less like a temple to a coming supermind and more like competent systems engineering: heterogeneous components, explicit interfaces, measured trade-offs, provenance that survives contact with procurement, and governance proportional to the damage a component can do. The monolith is not becoming sentient so much as becoming a dependency graph. That may be less thrilling than AGI. It is also where the invoices, outages, lawsuits, and actual advantage live.