Frontier Generalists or Specialised Models, Where Thought Leaders Think AI Is Heading

Narrative

Thoughtworks’ successive Technology Radars provide the clearest practitioner case for a durable specialised tier. Volume 31 put small language models in Trial in October 2024, citing privacy, latency and hosting constraints; Volume 33 argued that narrow, repetitive agent workflows are a good match for SLMs; Volume 34 reported that teams were already evaluating them for cheaper and faster summarisation and basic coding. This is a conditional operational judgement, not a forecast that small models displace frontier systems across difficult reasoning tasks.

The stronger pattern is that system design and evaluation determine whether a general model becomes useful in a domain. Bharani Subramaniam and Martin Fowler report that RAG satisfies most production cases they have encountered, with fine-tuning reserved for cases where task-specific behaviour cannot fit in retrieved context; their examples include a life-sciences corpus of 17,000 reports. DORA’s 2025 report similarly frames AI as an amplifier of organisational conditions, while Stack Overflow’s survey records high use alongside low trust, making human review, delivery practices and local context material constraints on deployment.

Vendor evidence supports specialisation but needs clear qualification. Databricks reports that its fine-tuned QuickFix Llama 8B beat its GPT-4o reference on accepted fixes and latency, and that prompt optimisation pushed an open model past a frontier model on its proprietary extraction benchmark at much lower serving cost. Those are vendor-run, task-specific evaluations rather than general capability results. Google Cloud and Databricks both now offer frontier APIs, open weights and tuning in the same platforms, which is practical evidence that enterprise architecture is becoming model-routing and context-management work rather than a one-time choice between general and specialised models.

Named voices (follow-up interview sweep): The evidence does not support a simple frontier-model-winner account. Satya Nadella says Microsoft expects enterprises to mix GPT-4o with Phi and other models for cost, quality and privacy, while his 2026 account makes private evaluations, reinforcement-learning environments and usage traces the durable firm-specific asset. Aidan Gomez makes a sharper commercial claim, that last year's model has little value, which shifts the question from owning a static model to controlling deployment, data and adaptation.

Open-weight advocates argue that the base model can be shared without eliminating application differentiation. Mark Zuckerberg expects a mixture of larger and smaller models, fine-tuning and tool-specific software, and says Meta's product work can stay differentiated even when Llama weights are available. Yann LeCun similarly treats open base models as a platform for many competing fine-tuned systems, while Arthur Mensch's Mistral position is that modifiable weights let regulated customers retain deployment control.

Several builders explicitly resist overconfident forecasts. Zuckerberg says he does not know the optimal division between scaling and specialised systems, and says Meta would reassess open release if capabilities changed qualitatively. Dario Amodei says Anthropic cannot perfectly separate the contribution of pre-training from post-training, while Cursor's Aman Sanger describes repository-specific training as an empirical question rather than a settled recipe. Nathan Lambert's account of DeepSeek adds a practical qualification: weights alone do not make a model fully reproducible when data and training code remain unavailable.