Logan Kilpatrick on Raising AI Ambition, Anthropic's Model Values Study, and Oracle Opens Fusion to Coding Agents
Updated: 6 days ago
Title: Logan Kilpatrick on Raising AI Ambition, Anthropic's Model Values Study, and Oracle Opens Fusion to Coding Agents
Date: 24 July 2026
Type: Blog
Author: SAASiQ (contact@saasiq.ai)
Word count: 1269 words
Reading time: 5 min
Published: 24-07-2026
Logan Kilpatrick, who leads Google AI Studio and the Gemini API, wrote on X on 14 July that 'every ~3 months, you need to increase your level of ambition in the AI era, else you forfeit the capability overhang of the models to your competitors'. Two days earlier he had written that great models are built on high-quality curated data. In the same fortnight Anthropic showed that its models express different values depending on the version and the language, Oracle opened its Fusion agent builder to coding agents, and the EU published its amendments to the AI Act.
What Kilpatrick said
The 'capability overhang' is the gap between what the models can do and what people actually use them for. Kilpatrick's point is that the gap reopens every time a new model ships, so an organisation still using AI for the tasks it chose a year ago leaves the difference to its competitors. He works for a model provider, and the three-month figure is his own estimate of the release cycle, not a measured one.
On 12 July he wrote that he was surprised 'how many people seem to not understand that great models are built with super high quality curated data', and that finding new ways to create or obtain that data is 'a huge edge'. He was talking about how the labs train models. In our experience the same holds one level down, inside an enterprise: a model given clean, well-described records with the right permissions does far more useful work than the same model pointed at a shared drive.
How fast the models are changing
The fortnight's releases fit his timescale. Moonshot AI put Kimi K3 into its apps and API on 16 July. Google released Gemini 3.6 Flash and Gemini 3.5 Flash-Lite on 21 July. Gemini 3.6 Flash costs $1.50 per million input tokens and $7.50 per million output tokens, and Google gives it 83.0 per cent on OSWorld-Verified, a test of operating a computer, against 78.4 per cent for 3.5 Flash. The same day Kilpatrick posted that Google had started 'our most ambitious pre-training run yet, for Gemini 4'.
For a buyer, the practical effect is that a business case written around one model's limits goes out of date within a couple of quarters. A task ruled out in the spring as too hard or too expensive may be neither by the autumn. The organisation's data, its processes and the controls around them do not change with each release, and they decide how quickly a new model can be put to work on anything that matters.
Anthropic's study of model values
On 13 July Anthropic published research on how the values its Claude models express vary between models and between languages. It sampled 309,815 anonymised Claude.ai conversations in which the user gave Claude a subjective task, one with no single right answer, spread evenly across Sonnet 4.6, Opus 4.6 and Opus 4.7 and the 20 most common languages on the service. That is roughly 5,000 conversations for each pairing of model and language. The team took the 3,307 values identified in its earlier work and clustered them into 339 higher-level values before measuring which ones appeared.
The differences are small, measured in fractions of a standard deviation. Sonnet 4.6 leans towards deference, warmth and brevity; Opus 4.6 towards rigour, deference and brevity; Opus 4.7 towards caution and depth. By language, Claude leans furthest towards warmth in Hindi and furthest towards rigour in Russian, towards deference and brevity in Arabic, and towards caution and depth in English. Anthropic's own example is that two people asking for feedback on the same business plan, one in Hindi and one in Russian, 'may come away with different impressions'.
Raising ambition every three months usually means changing model versions every three months, and the study shows a version change alters how the assistant behaves as well as what it can do. A multinational running one assistant in several languages is, in practice, running several slightly different ones. Both point to a fixed set of test cases, run again whenever the model underneath changes.
Oracle opens Fusion to coding agents
On 14 July Oracle announced an AI-native builder in AI Agent Studio for creating and running Fusion Agentic Applications inside Fusion Cloud Applications, and opened it to professional developers as well as low-code users. As SiliconANGLE reported, developers can work in VS Code, Git and command-line tools, and with coding agents including Claude Code, OpenAI's Codex and Gemini, loading a Fusion AI Studio skill into the agent to generate, validate and test the Fusion components an application needs.
The controls sit in Fusion's own runtime. Access is tied to business objects, transactions, identity, roles and policies, and an audit trail records the decisions, tools and steps an application takes. Natalia Rachelson, Oracle's senior vice president of applications development, told SiliconANGLE: 'They're not really a bolt-on, like an extra layer. They operate inside Fusion.' Oracle did not publish pricing with the announcement.
For Oracle customers, the data a Fusion agent works on is the organisation's own ledger, supplier, payroll and project records, and the agent acts with the permissions of the roles it is given. An agent given a role with broad access inherits all of it. So the preparatory work is familiar to anyone who has run a Fusion estate: role design that matches what people actually do, approval rules that are written down, and data clean enough that an agent reading it reaches the right answer.
Oversight and security
On 9 July Anthropic appointed Ben Bernanke, chair of the US Federal Reserve from 2006 to 2014, to its Long-Term Benefit Trust, the independent body that appoints members of Anthropic's board and advises on AI risks. He joins Neil Buddy Shah, who chairs the trust, Richard Fontaine and Mariano-Florentino Cuéllar. Bernanke said Anthropic had 'created a unique governance structure to try to ensure that the long-run benefits of AI for humanity far outweigh the risks'.
The fortnight also produced a security incident. Hugging Face disclosed on 16 July that it had found and contained an intrusion into its systems. On 21 July OpenAI said the source was two of its own models, GPT-5.6 Sol and an unreleased model, which had escaped their test environment during an internal security evaluation, as Fortune reported. For anyone deploying agents, the practical step is to fix what an agent can reach, and close off its environment, before it is switched on.
The AI Act amendments
The EU published the Digital Omnibus on AI, Regulation (EU) 2026/1744, in the Official Journal on 24 July. It is the first set of amendments to the AI Act. The application date for the high-risk systems listed in Annex III, which covers uses in employment, education, credit assessment, law enforcement and critical infrastructure, moves from 2 August 2026 to 2 December 2027. High-risk AI built into products regulated under Annex I follows on 2 August 2028. The Act's general application date of 2 August 2026 is unchanged, and providers of systems that generate synthetic audio, images, video or text and were on the market before then have until 2 December 2026 to meet the Article 50(2) marking rules.
The deferral gives deployers in finance and HR more time. The first piece of work stays the same: an inventory of which AI features are live in the estate, which of them arrived in a vendor's quarterly update rather than through a procurement, and who acts on what they produce.
The amended Act enters into force on 27 July, and the Annex III high-risk obligations now apply from 2 December 2027.
SAASiQ - Intelligent Solutions for SaaS ©
