AI news5
OpenAI's GPT-5.4 acts as near-autonomous chemist, completing a drug discovery project with lab-validated results across 10,080 reactions
Working with Molecule.one's Maria AI and its high-throughput laboratory, GPT-5.4 reviewed scientific literature, proposed that mild oxidants including TEMPO could improve Chan-Lam coupling for primary sulfonamides, designed experiments, and analysed the results — across 10,080 reactions run autonomously. Yields improved for 88% of boronic acids and 83% of sulfonamides tested; the mean yield rose from 16.6% to 25.2%. Human chemists validated the result at bench scale, with 11 of 14 substrate pairs showing higher yields and eight showing more than twofold improvement. The full process took three months, from March 4 to sharing with external experts on June 4.
openai.com ↗Why it matters
a frontier model that can propose a specific surprising hypothesis, design thousands of experiments, and have human chemists validate the result at bench scale is a qualitatively different scientific tool from a chatbot — improvements to sulfonamide chemistry (found in anticancer drugs, antimicrobials, and diuretics) make this a practically significant finding, not a benchmark score.
EPFL releases MeditronFO, the first fully open framework for building medical LLMs, making every stage of development publicly auditable
Researchers from EPFL's LiGHT lab released MeditronFO (Fully Open), a framework for building medical large language models that makes training data, code, training procedures, and evaluation methods all publicly available for independent review. The framework medicalises several open base models including OLMo, EuroLLM, and Apertus, building on the original Meditron released in 2023. Most medical AI systems in clinical use today are proprietary: their training data and decision-making processes are hidden, making independent verification virtually impossible.
actu.epfl.ch ↗Why it matters
the standard applied to AI in healthcare should match the standard applied to human clinicians. MeditronFO is the first systematic attempt to make every stage of medical LLM development auditable — the same kind of transparency that allows peer review and trust in medicine to function.
Google releases Gemma 4 as open source, with a 31B flagship ranking third on the open-source leaderboard and a 256K context window on a single GPU
Google released Gemma 4 with four models: a 31B dense flagship with 256K context (third on the Arena AI open-source leaderboard, runnable on a single H100), a 26B A4B MoE variant that activates only 3.8 billion of 25.2 billion total parameters per inference while outperforming comparable models, and E4B and E2B edge variants for mobile and embedded devices (E2B runs under 1.5GB). On the AIME2026 benchmark, Gemma 4 scored 89.2% compared to Gemma 3's 20.8%.
xix.ai ↗Why it matters
an open-source model at #3 on the leaderboard with a 256K context window running on a single GPU shifts what researchers, nonprofits, and developers can do without API costs or data leaving their own infrastructure.
Anthropic's Claude Fable 5 tops the Epoch Capabilities Index at 161 while GPT-5.6 reportedly targets a late-June launch
Anthropic's Claude Fable 5 scored 161 on the Epoch Capabilities Index this week, taking the top spot with dominant performances in math and software engineering above GPT-5.5 Pro. Multiple reports simultaneously placed OpenAI's GPT-5.6 on a late-June timeline, with OpenAI chief scientist Jakub Pachocki reportedly describing it as "a meaningful leap." Reports cited a 1.5 million token context window for GPT-5.6, with performance reportedly exceeding Claude Fable 5 on some benchmarks at approximately one-third the cost per token.
cryptobriefing.com ↗Why it matters
the frontier model cycle is now measured in weeks. Organisations building AI strategies around "current best model" need evaluation pipelines that can keep pace, and cost-per-token differences at this scale affect what's financially viable to deploy.
Thirty mathematicians gather at Harvard to formally grade AI on unpublished research problems, and seven of ten pass
In the second round of First Proof, a mathematician-led benchmarking project, 30 mathematicians gathered at Harvard to grade four AI systems on ten genuine research-level problems that had been privately solved but never published. Seven of ten problems had at least one correct AI solution; some solutions were described as "flawless" and one impressed referees by using a different strategy to the human solution. The initiative follows OpenAI's May announcement that an internal model disproved Paul Erdos's 80-year-old planar unit distance conjecture, which Fields Medallist Timothy Gowers said deserved publication in Annals of Mathematics "without any hesitation."
iol.co.za ↗Why it matters
First Proof closes the training-data loophole that lets AI appear capable by regurgitating memorised content. Seven of ten passes on unpublished research-level problems demonstrates genuine mathematical reasoning, not retrieval; the three failures mark the current ceiling.
AI in the nonprofit sector4
Anthropic launches Claude Corps with $150M, placing AI fellows full-time inside nonprofits at $85K each
Anthropic announced Claude Corps, a $150 million fellowship program to train early-career workers in AI and place them full-time with US nonprofits for 12 months. The program starts with 100 fellows and is designed to scale to 1,000 fellows at up to 400 host nonprofits. CodePath serves as employer of record and leads programming; each fellow receives $85,000 in salary plus benefits and Claude API access, while each host nonprofit receives a $10,000 implementation grant. Applications open on a rolling basis through July 17, 2026, with fellows starting October 19.
thenonprofittimes.com ↗Why it matters
placing paid AI practitioners inside nonprofits for a full year is a fundamentally different model from one-day workshops or tool subscriptions. At scale, 1,000 fellows across 400 nonprofits would represent the largest AI capacity-building deployment in the nonprofit sector.
OpenAI Foundation commits $50M to a second People-First AI Fund round, with applications open to US nonprofits through July 15
The OpenAI Foundation announced a new $50 million commitment for 2026, building on a 2025 round that distributed $40.5 million in unrestricted grants to 208 nonprofits. The 2026 fund focuses on three areas: community support services (legal aid, public benefits, disability access), community arts and cultural organisations, and community journalism and media. Grants are unrestricted and up to 10% of an organisation's annual operating budget. Eligible organisations must be US-based 501(c)(3)s with budgets between $500,000 and $10 million. Applications close July 15, 2026.
opportunitydesk.org ↗Why it matters
two consecutive $50M rounds targeting direct-service community organisations signals that the OpenAI Foundation is positioning itself as a sustained philanthropic actor in the sector rather than a one-time gesture, and the unrestricted grant structure gives nonprofits genuine flexibility to use it where it matters.
Indonesia deploys AI-backed social protection system to cover 35 million families, cutting registration time from 200 days to minutes
Indonesia is rolling out Perlinsos Digital, an AI-backed social protection platform that allows 35 million social aid recipients to register using their 16-digit national ID and facial verification. The system integrates data from eight ministries and is projected to save between $10 billion and $15 billion by eliminating leakages and ensuring aid reaches eligible recipients. Registration time drops from up to 200 days to minutes; costs fall from Rp150,000 per applicant to near zero. A pilot is underway in 42 cities and regencies with nearly 370,000 residents already registered; national launch is planned for October or November 2026.
jakartaglobe.id ↗Why it matters
deploying AI to clean up social welfare eligibility at the scale of 35 million families in a developing-economy context is one of the largest public-sector AI deployments in social protection globally, and the cost-saving projections, if realised, would redirect substantial government resources toward actual recipients rather than administration.
More than 200 civil society organisations call for an immediate halt to AI in military kill chains, warning of humanitarian law violations
A joint statement signed by more than 200 civil society groups and advocates, including the World Council of Churches and Access Now, called for an immediate halt to the use of AI systems in military "kill chains" on June 16. The statement warns that AI-accelerated warfare risks facilitating violations of international criminal, human rights, and humanitarian law by compressing decision cycles and removing meaningful human accountability for targeting decisions. The statement was released simultaneously at a Geneva event and by signatories globally.
jurist.org ↗Why it matters
more than 200 organisations signing a unified statement represents the most coordinated civil society position on AI and warfare to date. The nonprofit sector is staking out its clearest collective line on where AI should not go, at the same moment that AI capacity-building investment is accelerating.