Anthropic builds AI Assistant for GOV.UK for Job Help

UK Takes $1M From Meta While Claude Enters GOV.UK Services Anthropic has secured a rare type of AI contract that can significantly impact the operations of a country. The UK government picked Anthropic to build a Claude assistant inside GOV.UK, starting with job and training help. The proposal aims to provide the correct form, benefits, and training path in response to a single question. The risk is that individuals may miss benefit deadlines, support, job opportunities, submit incorrect forms, or receive misleading guidance if the assistant malfunctions.

Here’s what the pilot actually includes: Scope: A Claude-powered helper inside GOV.UK aimed first at employment services.

• Control: Anthropic says users control what the tool remembers and can opt out of memory.
• Testing: The rollout follows the government’s 'Scan, Pilot, Scale' approach with evaluation support tied to the UK AI Safety Institute.
• Risk: A designated assistant for job-related queries may still provide incorrect answers with confidence or inconsistent guidance to various users.

Politics surrounding this issue are intensifying as well because this comes right after Anthropic’s CEO warned the public to 'wake up to the risks,' while ministers also took $1M (£728k) from Meta to fund four British AI experts coordinated by the Alan Turing Institute. That money landed during a consultation on a possible under-16 social media ban that could hit Instagram. The UK wants AI that helps citizens move faster but it also keeps building its future with the same companies it may need to police. OpenAI EU Blueprint 2.0: Helping Small Firms Adopt AI OpenAI has published its EU Economic Blueprint 2.0 and teamed up with Booking.com on an SME training push across France, Germany, Italy, Poland, Ireland, and the UK, highlighting a diverse range of European countries.

Europe has sufficient AI access but lacks consistent day-to-day usage. OpenAI says power users consume about 7× more ‘thinking capabilities’ than typical users, and 9 EU countries still sit below the global average. Here is what OpenAI is proposing: Training: OpenAI and Booking.com say they will help train 20,000 SMEs through OpenAI Academy with online and in-person sessions.

• Grant: It has set aside €500,000 for NGO research on youth safety and well-being linked to AI use.
• Gap: The problem is 'capability overhang,' meaning the tools exist but many people and countries barely push them.
• Index: OpenAI proposes an AI Adoption Index to track depth of use across countries and sectors.

OpenAI highlights a significant gap in business adoption: only 17% of small businesses utilize AI, whereas 55% of large firms do. This looks like OpenAI trying to win Europe through spreadsheets, training, partnerships, and policy language rather than just bigger models. When the company uses its own metrics to track progress, Europe may confuse product activity with productivity, making measurement challenging. For Europe to adopt AI without relying solely on it, a common standard is essential, even in OpenAI's absence.

Google

DeepMind’s AlphaGenome Targets 98% ‘Dark DNA’ Google DeepMind’s AlphaGenome targets the 98% of DNA that does not code for proteins, where many disease-linked variants sit. DeepMind opened access through a non-commercial API for researchers. The idea is to feed it up to 1 million and rank which single-letter DNA changes actually alter gene regulation and predict changes in gene activity, splicing, etc. It could cut months off the hunt for which mutation matters, but a confident score can look like a conclusion when it is still a prediction.

Here’s what the evidence suggests: Scope: Predicts 5,930 human and 1,128 mouse genome tracks across 11 modalities.

• Window: Reads up to 1 million DNA letters at once so it can model far away regulatory effects.
• Users: Built for geneticists and disease teams who keep running into variants that do not explain the biology.
• Limits: Docs warn it is only good for limited predictions, while DeepMind and outside scientists say it is not a clinic-ready oracle.

This will likely help researchers narrow down which DNA changes deserve expensive follow-up work and that alone can save months. The danger is the same one we saw after the hype with AlphaFold, an AI system that predicts the 3D shape of proteins from their amino acid sequence, where people start treating a strong prediction like a finding. If DeepMind wants real credibility, the next wave should be boring and specific, with outside labs showing repeatable wins and clear failure cases.

• Human Flow: AI Hiring Copilot for Startups Human Flow is a lightweight ATS that helps you move faster from “inbox of resumes” to “clear next steps.” You can use it to score candidates against a role, generate interview briefs, and turn interview transcripts into structured notes and follow-ups.

It’s built for small teams that want a simple loop: screen, prep, interview, decide. Core functions (and how to use them): Resume scoring: Upload a resume and match it to a job description to get strengths, gaps, and a quick fit summary you can share with the team. Interview briefs: Generate a short prep doc with targeted questions based on the candidate’s background and the role requirements. Transcript analysis: Paste an interview transcript to extract claims, examples, red flags, and follow-up questions you should ask next round.

Consistent comparisons: Use the same role requirements across your shortlist so every candidate is judged on the same checklist, not who interviewed them. Ask-anything search: Type questions like “Who has shipped production ML pipelines?” to quickly find candidates that match specific needs. Try this yourself: Pick one open role and 3 recent resumes. Paste your job requirements, then run each resume through Human Flow.

For each candidate, save: the top 3 strengths, the top 2 missing requirements and 5 interview questions you’ll reuse for everyone. Now you’ve got a repeatable screening template you can use for every future batch.