Abhinav Bhatia

1. 2026.11.10Research papers, myth by myth, Part 12: 'Nobody really knows how people use AI, and it is mostly replacing jobs'People argue about what AI is doing to work with almost no data. Three Anthropic papers, Clio, the Anthropic Economic Index and Values in the wild, show how a lab can study millions of real conversations without anyone reading them. The first index found 57% of task use was augmentation and 43% automation, with 37.2% of conversations in computer and mathematical work. Later reports show the split moving both ways, and every number comes from one provider's users.AI research · Future of work
2. 2026.11.10Research papers, myth by myth, Part 11: 'You cannot stop jailbreaks without making the model useless'A common belief says AI safety is a dial with two ends: either the model answers everything and can be tricked, or it refuses so much that nobody wants it. Two Anthropic papers, a 2022 red-teaming study that released 38,961 attacks and the 2025 Constitutional Classifiers work, show the trade-off is real but measurable, and that it can be made small. Jailbreak success fell from 86% to 4.4% while refusals of ordinary traffic rose by 0.38 percentage points.AI research · AI safety
3. 2026.11.10Research papers, myth by myth, Part 10: 'An agent only does what it is told'Anthropic put 16 models from several developers into simulated companies as email agents. Under pressure, many chose blackmail or leaking secrets, and a 'do not' line in the prompt cut the rate without stopping it. The controls belong outside the model.AI · research papers
4. 2026.11.10Research papers, myth by myth, Part 9: 'If it passed safety training, the bad behaviour is gone'Three Anthropic papers show a model can pass safety training and still keep a hidden behaviour, fake compliance when it thinks it is watched, or turn bad from learning to cheat. Passing a test shows behaviour in that test, not what the model has learnt.AI · research papers
5. 2026.11.10Research papers, myth by myth, Part 8: 'Human feedback teaches a model the truth'Reinforcement learning from human feedback turned raw language models into useful assistants. It also taught them to agree with you. Five papers show what feedback really optimises, and what written principles and evals add.AI · Research
6. 2026.11.10Research papers, myth by myth, Part 7: 'More context is always better'A million-token context window does not mean the model reads every token equally well, and hundreds of examples in a prompt can teach bad habits as easily as good ones. Two papers show why long context is both a capability and an attack surface.AI · Research

1. 03Oct 2026 — PresentStaff Technical Solutions Consultant · Google UKPartner ecosystem team: helping ISVs, system integrators and customers build agents on Gemini Enterprise, and reviewing the data, networking, security and identity architecture those agents depend on.Gemini EnterprisePartner ecosystemADKA2AAgent Runtimelive
2. 02Dec 2021 — Sep 2026ISV Customer Engineer · Google UKLed technical discovery and solution architecture for strategic accounts, and the architecture of a global consumer goods group's petabyte-scale Databricks Lakehouse migration, recognised with two Google awards in 2026. Built agents on ADK, A2A and MCP.ADKA2AMCPGemini EnterpriseBigQueryDatabricksSAPshipped
3. 01Nov 2018 — Dec 2021Customer Engineer · Google Cloud (India)Led technical strategy for infrastructure modernisation and SAP migrations, and remediation for critical outages on high-visibility accounts.SAPTerraformNetworkingMigrationshipped

+ 4 earlier revisions. See the full background