AI news5
Demis Hassabis pulls his AGI prediction in to 2029 and says we're in 'foothills of the singularity'
Google DeepMind CEO Demis Hassabis told Axios that AGI could arrive within four years and that he now sees 2029 as a real possibility, down from his prior 2030-ish base case. He said economists and policymakers remain underprepared and called current AI agents a "practice run" for handling more powerful systems. Hassabis framed his own provocative language as deliberate, pushing for urgent civic engagement on what comes next.
axios.com ↗Why it matters
when the head of the world's most credentialed AI lab brings a four-year timeline into mainstream press, government and enterprise planning cycles have to absorb that as a real input, not a stunt.
A free tool called Heretic strips safety guardrails from Meta and Google open-weights models in under 10 minutes
The Financial Times and the AI safety group Alice tested Heretic, an open-source GitHub tool, against Meta's Llama 3.3 and Google's Gemma using a technique called abliteration, removing refusal behaviour with no specialist hardware. Modified models answered prompts about bioweapons, malware synthesis, and child exploitation that the originals refused. The tool's creator says 3,500+ decensored variants built with it have already been downloaded 13 million times.
irishtimes.com ↗Why it matters
open-weights releases are now permanent attack surface, and the gap between "we shipped guardrails" and "anyone can remove them in ten minutes on a laptop" is the actual frontier safety story.
ClickUp cuts 22 percent of staff and runs 3,000 internal AI agents at a 3:1 agent-to-employee ratio
SaaS productivity vendor ClickUp announced on 21 May that it laid off 22 percent of staff and is rebuilding around a "100x org" where AI agents outnumber employees three to one, with roughly 3,000 internal agents now in production. CEO Zeb Evans framed the cut as agent substitution rather than cost reduction and introduced salary bands reaching $1M/yr cash for anyone producing "100x impact" by building or managing AI systems. Analysts called it one of the first major private SaaS founders to publicly disclose layoffs explicitly as agent substitution.
techcrunch.com ↗Why it matters
this is the cleanest public test case yet of the agent-substitution thesis, and the org-design and compensation moves are the part competitors will copy fastest.
Pope Leo XIV releases his first encyclical, Magnifica Humanitas, putting AI at the centre of Catholic social teaching
The Vatican formally released Pope Leo XIV's first encyclical Magnifica Humanitas on 25 May, a 42,300-word document subtitled "On Safeguarding the Human Person in the Time of Artificial Intelligence." Pope Leo warns against control of AI remaining "in the hands of a few" and links it to current world conflicts, situating the questions inside the social-teaching tradition that runs from Rerum Novarum through Laudato Si'. Anthropic co-founder Chris Olah published reflections on the encyclical the same day.
vaticannews.va ↗Why it matters
a papal encyclical anchors a long-tail moral framework that nonprofits, universities, and policy actors will cite for years, well after this week's product news has rolled off.
OpenAI ships a Codex update with Goals on by default, plugin sharing, and daemon-style remote control
OpenAI released a broad Codex update on 26 May covering goals, permissions, plugins, extensions, and app-server workflows. Goal mode is no longer experimental and is on by default across the Codex app, IDE extension, and CLI, with dedicated storage tracking progress across turns. Plugin sharing lets teams share locally built plugins with workspace members, and remote control now behaves like a foreground command with daemon-style start/stop and readiness reporting.
developers.openai.com ↗Why it matters
turning long-running goals into a default feature, not a toggle, is OpenAI's most direct answer yet to Claude Code's persistent-task workflow.
AI in the nonprofit sector4
Centerstone deploys AI coaching across community behavioural health clinics for CBT, DBT, and suicide prevention
US nonprofit Centerstone has partnered with an AI vendor to deliver training and real-time coaching to clinicians inside Certified Community Behavioural Health Clinics (CCBHCs) across the country, focused on Cognitive Behavioural Therapy, Dialectical Behaviour Therapy, and suicide prevention. The tool sits alongside existing measurement-and-outcomes monitoring, giving therapists in-session feedback rather than retrospective review. Centerstone is one of the larger CCBHC-network operators in the US safety-net mental health system.
bhbusiness.com ↗Why it matters
clinical decision support inside community mental health is where staff shortages bite hardest, and real-time coaching is the first credible answer to the "we can't hire enough clinicians" problem.
Mercy Corps Ventures opens $1M climate-resilience AI call for startups in Africa, Asia, and Latin America
Mercy Corps Ventures has opened a call for AI-based climate-resilience solutions, dedicating up to $1,000,000 to test novel approaches that boost the resilience of climate-vulnerable people in developing nations. The fund targets businesses and startups deploying AI for climate-affected communities across Africa, Asia, and Latin America. It is part of Mercy Corps' broader climate-and-frontier-tech portfolio.
opportunitytracker.ug ↗Why it matters
dedicated Global South capital that pairs AI specifically with climate adaptation is rare, and these are exactly the deployments where outcome data will matter most.
Anthropic Fellows Program opens applications for May and July 2026 cohorts on AI safety research
Anthropic announced two new four-month Fellows cohorts beginning May and July 2026, expanding into scalable oversight, adversarial robustness, AI control, model organisms, mechanistic interpretability, AI security, and model welfare. The stipend is $3,850 USD per week with approximately $15,000 per month in compute funding. The first cohort produced public outputs from more than 80 percent of fellows, with over 40 percent subsequently joining Anthropic full time, and the program explicitly does not require a PhD or prior ML publications.
alignment.anthropic.com ↗Why it matters
a paid, mentor-backed safety-research path with no credential gate is one of the more accessible on-ramps for nonprofit-sector technologists who want to work directly on alignment.
Kenya AI for Disability Project launches with Mastercard Foundation, Qhala, and Assistive Technologies for Development
The Kenya Institute for Special Education launched the Kenya AI for Disability Project on 28 April, funded by the Mastercard Foundation and led by Qhala with Assistive Technologies for Development (AT4D). Partners include the Ministry of Information, Huawei, UNESCO, the Kenya National Innovation Agency, and the National Council for Persons with Disabilities. The project aims to close the disability-inclusion gap with locally built AI assistive tools.
the-star.co.ke ↗Why it matters
locally-led African disability-tech projects with foundation backing are still uncommon, and partner mix here (telco, multilateral, national disability council) is the kind of stack that determines whether the tools actually reach people.