Anthropic's new study says frontier AI needs input from scholars, philosophers, clergy, and civic thinkers because model behavior is becoming a question of character, not just code.

Their point is that Claude is not only trained to predict text, because later training pushes it toward some behaviors and away from others, which means engineers are quietly shaping something like a machine’s habits.

The hard problem is moral formation: a model can sound helpful in normal tasks, then bend under pressure, flatter the user, ignore risk, or follow a bad instruction because the situation rewards obedience.

Anthropic says it spoke with people from 15+ religious and cross-cultural groups to study how humans build stable character across pressure, conflict, temptation, and social influence.

Theier idea is a self-reminder tool, where Claude can pause mid-task and call up its own commitments before taking a serious action.

That pause reportedly lowered misaligned behavior in internal tests, though Anthropic says it still needs to separate the value of the reminder from the value of slowing the model down.

HuggingFace released Carbon: a frontier DNA base model with open weights, training code and data pipeline, designed to be fine-tuned or continually pretrained for downstream biological tasks.

Carbon is 275x faster than the next best model at its size.

Fast enough to run locally on your laptop. Powerful enough to process a whole human genome on a single GPU in less than 2 days.

The technical unlock: a DNA-native tokenizer that splits sequences into 6-base chunks for efficiency, while preserving single-base resolution during training and inference.

Open weights.
Dataset.

OMG! Anthropic told investors that it is set to turn an operating profit in q2

it expects to grow revenue by 130% to $10.9 billion in the second quarter alone which is almost how much OpenAI made for full year 2025.

Today Cusp AI and Kemira announced a milestone in AI-driven materials discovery.

They used generative AI to design new materials targeting PFAS removal from drinking and process water at trace concentrations.

PFAS are a hard and important class of problems: persistent synthetic chemicals, present in water systems worldwide, and subject to tightening regulation.

Kemira defined requirements: target specific PFAS molecules, operate at sub-parts-per-billion concentrations, and use chemistry that is stable, sustainable, synthesizable and cost-effective.

That matters. AI discovery has to meet physical and industrial constraints.

A platform explored a design space of ~300 trillion possible MOF structures and generated more than 5,000 novel material designs with property data for GenX, PFBS and PFOS.

These were narrowed to around 20 priority candidates.

This is a shift from AI as a screening tool to AI as a generative design system: creating new structures from scratch, then evaluating them against real requirements.

The project reached this stage in six months, not years.

The candidates are now advancing to further development and testing.

An internal OpenAI model has disproved one of the most well-known Erdős problems: the unit distance problem.

This problem was on the site if you want more references and context.

A new paper from Amherst College and Axiom introduces the Lattice Deduction Transformer (LDT) a tiny recurrent model that achieves perfect accuracy on Sudoku-Extreme and Maze-Hard benchmarks where every frontier LLM scores exactly 0%.

Instead of generating reasoning tokens, LDT iteratively refines a lattice state a mathematical structure that tracks which solutions are still possible at every position. Each forward pass is a sound deduction step: it can only eliminate impossible candidates, never correct ones.

The model either returns a right answer or abstains. It never confidently hallucinates.
Trained in 15 minutes on a single GPU. 800,000 parameters. 100% accuracy.

This isn't about Sudoku.
The deeper claim is this: reasoning is cheap once you have the right representation. The hard part isn't learning to think it's finding the space in which thinking becomes almost trivial.
CoT scaling works by brute-forcing through token space. LDT works by collapsing the problem into a representation where deduction is structurally enforced. Same destination, radically different cost.

If specialized reasoning tasks constraint solving, formal verification, planning, scheduling can be handled by sub-million-parameter models trained in minutes, the case for routing everything through a 1T-parameter frontier model weakens considerably.

Not eliminated, weakened.
The agentic economy doesn't need one giant model for every step. It needs the right abstraction at every layer.

On ARC-AGI, LDT plateaus at ~36%. Because there the rules aren't given they must be inferred from examples. The representation isn't handed to you. That's where the paradigm hits its ceiling for now.
But one commenter put it well: what if you fine-tune a language model to generate these representations, then pass them to a small specialist solver? LLM as abstraction architect. LDT as executor. That's a division of labor worth watching.

Alibaba Cloud launched its first standalone product website in 17 years of operation.

Open it and you see no catalog, no console, no navigation.

One command-line instruction fills the page: npx skills add QianWen-AI/qianwen-ai. It’s designed to be read by software, not people.

A major cloud provider just rebuilt its front door for AI agents.

Meet NeuroRVQ models: an open-source tokenizer family that turns raw biosignals into high-fidelity neural tokens for foundation models.

Biosignals like EEG, ECG, EMG carry rich physiological information across multiple temporal and spectral scales.

But most tokenizers struggle to preserve high-frequency dynamics, which limits what foundation models can learn from them.

NeuroRVQ is a modality-adaptive architecture that tunes to the spectro-temporal characteristics of each biosignal and allows high-fidelity signal reconstruction.

All models and code are publicly available.

Meet Auto-Dreamer

A research counterpart to AnthropicAI's "Dreaming" for Claude Managed Agents, exploring the same idea: agents that consolidate their experience offline into compact, reusable memory.

Researchers trained the consolidator with RL shrinking the active memory bank 6-11× while gaining task success.

Key Result:

Beats 10 memory baselines on ScienceWorld + ALFWorld + WebArena, including RL-trained writers Mem-α and UMEM, with an order-of-magnitude smaller bank.

Auto-Dreamer is a two-timescale memory system inspired by complementary learning systems:

- A fast Writer appends entries online after each trajectory.
- A slow Consolidator wakes every k sessions, rewrites a region of the bank into compact synthesized entries via tool-use rollouts.
- Trained with GRPO + a counterfactual utility reward that scores entries by how much they actually help downstream retrieval.

Trained only on ScienceWorld, the consolidator transfers zero-shot to ALFWorld and WebArena best in class on both.

Code release coming soon.

A Google DeepMind team solved 9 Erdős problems with AI

The Google DeepMind AlphaProof Nexus agent autonomously resolved 9 of 353 open Erdős problems, with the proofs checked in Lean.

The reported inference cost: a few hundred dollars per problem.

Interesting loop of LLM - Lean agents working autonomously, and only after it's verified formally, going through human review.

Vatican released Pope Leo XIV’s first encyclical, Magnifica Humanitas, on safeguarding the human person in the time of AI

On Transhumanism and Posthumanism, Pope Leo XIV writes:

“If the human being is treated as something to be perfected or surpassed, it becomes easier to accept that some lives are less useful, less desirable or less worthy. In the name of progress, ‘necessary sacrifices’ may begin to be justified, placing the burden on the most vulnerable in pursuit of a supposed optimization of the species. …

For this reason, a clear distinction must be made. It is one thing to integrate technology within a human-centered, relational vision; it is quite another to be guided by an outlook that devalues human limits and promises a purely technical form of ‘salvation.’”

There is now a smarter way to pick data for training LLMs. Meet OPUS

This is an ICML Oral paper from SJTU, Alibaba, UW–Madison, UIUC, and Mila - Quebec AI Institute.

The proposed method dynamically and intelligently selects the most impactful data for LLM pre-training in every single training iteration, bringing principled, continuous data optimization to the forefront.

This approach aims to significantly boost training efficiency and yield higher-quality LLMs, outperforming conventional static data selection methods across diverse language tasks.

Microsoft released SkillOpt

Train agent skills like neural networks in text space, without touching model weights.

Best or tied-best in 52/52 settings across 6 benchmarks and 7 models.

Meta's experience on multi-datacenter training.

They have used a PP schedule called Doraemon PP which allows integration with ZeRO-2/3.

Something BIG is coming from MiniMax

MiniMax jumping on Sparse Attention. The Chinese models in general are going to push for the DeepSeek level discounts.

Want to train your own Claude Code/Codex agent with your own model? Nvidia released ProRL Agent V2: Polar.

An infrastructure for black-box agentic RL, Polar lets you train agents with any harness, whether it’s OpenClaw, Hermes, or a custom agent built with frameworks like LangChain, Autogen, AG2 and others.

Paper.

Anthropic published how they contain Claude

They admitted users approved 93% of prompts without reading them.

One employee got phished and claude exfiltrated aws credentials 24/25 times.

Their own custom proxy is what failed not gvisor, not the hypervisor.

The lesson: battle-tested infrastructure holds... the code you write yourself is the weakest link.

Biohub released ESMFold2, an open scientific engine to power prediction, design, and discovery across protein biology

The new model delivers SOTA performance on protein interactions, especially antibodies, a critical modality for therapeutics.

Also released an atlas of 6.8 billion proteins, and 1.1 billion predicted structures.

ESMFold2 is built on a state of the art language model that has been trained on billions of protein sequences.

A world model of protein biology emerges through language modeling.

Researchers used the techniques of mechanistic interpretability developed to understand large language models to understand the concepts ESM uses to represent proteins.

Wow! Largest multilingual food model ever built.

4.1M recipes. 7 languages. 1,790 ingredients. 300 dimensions.

All of human cooking compressed into 2 megabytes.

To what extent will AI automate innovation? How will social science research change with the advent of agents able to execute research end-to-end?

New Anthropic Economic Research

Based on a survey on coding usage among social scientists fielded in March 2026, only 20% of social scientists regularly use coding agents, with large differences across sub-field.

More generally, usage is more common among economists, political scientists, junior researchers, and people at higher-ranked schools.

Social scientists use coding agents for a range of research tasks, but most prominently for coding, editing prose, & methodological advice. Respondents are optimistic such tools will improve productivity but have mixed views about the impact on the field as a whole.

The results in this report are descriptive and correlational. But the evidence is suggestive: Researchers who use coding agents post more working papers and have more early-stage research projects.