Tuesday, July 7, 2026

AI Legal Startup Norm Valued at $1.2 Billion

Norm, an AI-powered legal startup, reached a $1.2 billion valuation in a recent funding round. The round highlights growing investor interest in AI for legal services.

Meta debuted new AI image-generation model in chatbot and Instagram

The model is integrated into Meta's chatbot and Instagram, enabling users to generate images. No specific model name or capabilities were disclosed in the Bloomberg report.

First American autonomous ground vehicles deployed in Ukraine

Forterra has deployed more than 100 autonomous ground vehicles to Ukraine, the first such American vehicles in combat. The vehicles were developed by Forterra, a US-based autonomous systems company.

Verification loop boosts DeepSeek to match Opus at 1/7 cost

A verification loop increased DeepSeek's coding agent intelligence by 4x, achieving performance comparable to Opus at one-seventh the cost. The technique adds a self-verification step to improve reasoning without additional parameters.

Scaling Properties of Continuous Diffusion Spoken Language Models

Apple ML Research paper introduces the phoneme Jensen-Shannon divergence (pJSD) metric to quantify linguistic quality of spoken language models. Scaling continuous diffusion SLMs to 16B parameters with tens of millions of hours of data enables emotive, prosodic, multi-speaker multilingual speech, though long-form coherence remains challenging.

AWS details multi-turn RL infrastructure for Amazon Nova on SageMaker HyperPod

The post covers training agents that query databases, call APIs, and recover from mid-process failures. It describes infrastructure setup on SageMaker HyperPod for multi-step RL workflows.

Google Antigravity adds Predicting the Past Skill for Greek and Latin texts

A Single Neuron Is Sufficient to Bypass Safety Alignment in Large Language Models

Apple ML Research demonstrates that targeting a single neuron in either of two distinct systems—refusal neurons (which gate expression) or concept neurons (which encode knowledge)—can bypass safety alignment in LLMs. The paper details both directions of bypass.

Sakana AI launches Namazu-powered translation tool Sakana Translate

Sakana Translate handles bidirectional translation across Japanese, English, and Chinese. It's available as a free web app with three modes: Translate, Proofread, and Ask.

FlowEval: reference-based evaluation for generated UIs

FlowEval is a reference-based evaluation framework for assessing LLM-generated user interfaces. It addresses the difficulty of reliably evaluating visual and interaction design. Existing methods rely on human experts, making automation challenging.

MIRA: Multiplayer world model trained on Rocket League released

MIRA is a 5B parameter world model trained on 10,000 hours of synthetic Rocket League data, running at 20 fps for 4 players on a single B200. A playable online demo and in-depth details have been released.

Japan launches Noetra consortium to develop AI model, deploy 10M robots by 2040

The government-backed consortium includes SoftBank, Sony, Honda, and NEC. It aims to create a homegrown AI model and deploy 10 million robots by 2040.

ECB asks banks for plans to address AI cybersecurity threats

The European Central Bank's top supervisor Claudia Buch sent a letter to bank CEOs requesting action plans for AI cybersecurity risks by end of October. The move reflects growing regulatory focus on AI-related threats in the financial sector.

Depth Anything 3 predicts spatially consistent geometry from arbitrary views

Apple researchers tame text-to-sounding video generation with modality conditioning

The paper addresses two challenges: weak text conditioning and misalignment between audio and video modalities. It proposes a framework integrating cross-modal attention and joint conditioning to improve synchronization.

FLUX.2-9B-klein text encoder pruned 38%, fits 16GB VRAM

Pruning removes 38% of text encoder parameters (8.2B to 5.10B). Entire fp8 pipeline fits in 16GB VRAM for 768x768 without offloading.

Hugging Face introduces Tau, an educational coding agent

Tau is a small, readable coding agent written in Python, supporting Hugging Face Inference Providers, GLM 5.2, Minimax, and Kimi. It features a three-layer architecture, durable sessions, and is designed as an educational tool for learning agentic coding.

SWE-Marathon: Evaluating Coding Agents at Billion-Token Scale

SWE-Marathon includes 20 project-scale tasks covering product clones, library rewrites, and ML engineering, requiring agents to run for tens to hundreds of millions of tokens. The benchmark emphasizes the need for computer-use verifiers in full-stack evaluations.

Brahe launches as AI-first law firm

Founded by legal AI expert Antti Innanen, who previously created the agentic experimental firm Lavern, Brahe aims to provide AI-driven legal services. The firm is built on an AI-first approach, integrating AI into all aspects of legal practice.

Study finds LLMs fail to simulate human preferences

New study tested LLMs against thousands of real users and found they systematically fail to capture human preference distributions. The results cast doubt on the growing trend of replacing human feedback with synthetic user data.

Fastcase and Alexi head to court over legal AI data licensing

A federal judge in Washington, D.C. is hearing arguments in the Fastcase v. Alexi Technologies case regarding the use of licensed caselaw data. The outcome could establish legal precedents for how AI companies access and utilize proprietary legal datasets.

Nations Deploy AI for Strategic Priorities

Nations are investing in domestic AI infrastructure to advance economies, protect data, and seize opportunities in transportation, healthcare, and other sectors. The blog from NVIDIA's Calista Redmond highlights AI as a key technology for national priorities.

Together AI optimizes inference latency with ATLAS and NVIDIA Blackwell

Understanding Annotator Safety Policy with Interpretability

Apple ML Research's paper analyzes how annotator disagreement on safety policies can stem from operational failures or policy ambiguity. It uses interpretability methods to understand and improve annotation consistency.

Google's elastic training recovers TPUs from failure in seconds

Google's elastic training via Pathways lets multi-node TPU training survive individual machine failure, recovering in seconds. The capability, part of MaxText, was demonstrated by terminating a TPU mid-training with full recovery.

LlamaIndex releases legal-kb: agentic retrieval for legal documents

The legal-kb app uses LlamaIndex Index v2 with tools for retrieve, find, read, and grep. It is an open-source reference application for legal document knowledge bases.

Apple proposes Path-Constrained MoE for efficient inference

Apple researchers introduce Path-Constrained Mixture-of-Experts, which views expert token routing as paths across layers to reduce computation. The method addresses the exponential explosion of N^L possible paths while maintaining model quality.

Inside Zipline's Autonomous System: 140M Miles, Zero Incidents

Zipline's autonomous delivery system has flown 140 million miles with zero safety incidents, according to co-founder Keller Rinaudo Cliffton. He notes the drone itself is only 15% of the overall system, with the rest being software and infrastructure.

Apple proposes LensVLM for efficient visual text representation

LensVLM selectively expands context for compressed visual tokens in VLMs, enabling text processing as rendered images without long token sequences. The method varies rendering resolution to balance efficiency and accuracy.

HIVE raises $15M pre-Series A for physical AI in industrial machines

The London-based startup raised $15M in pre-Series A funding to build an intelligence layer for industrial machines. HIVE enhances existing equipment with sensors and AI for automation and teleoperation.

MiniCPM-5: 1B on-device AI model built for agentic tool use

The 1B parameter model features a 128K context window and is optimized for agentic tool use on mobile devices. It claims superior token efficiency compared to larger reasoning models.

Apple proposes Segmental Attention Decoding for long-form acoustic encodings

The method addresses the incompatibility of attention-based encoder-decoder models with long acoustic sequences, enabling handling of absolute frame positions. This extends AED models to long-form speech without failing generalization.

DeepL CEO Jarek Kutylowski explains rise of specialized AI models

Jarek Kutylowski argues purpose-built specialized AI models can deliver better accuracy, lower latency, and reduced costs compared to frontier general-purpose systems. He explains why these models are increasingly challenging larger labs despite their scale and resources.

Automatically redact PII in images with Amazon Nova

AWS introduces a new feature using Amazon Nova to automatically detect and redact personally identifiable information (PII) in images. The guide covers setup, configuration, and best practices for integration.

Apple launches Weblica for scalable web agent training

Apple ML Research introduces Weblica, a platform for scalable and reproducible training environments for visual web agents. It supports both offline trajectories for supervised fine-tuning and simulated environments for reinforcement learning.

Liquid AI open-sources Antidoom to fix reasoning model doom loops

Liquid AI released Antidoom, an open-source Final Token Preference Optimization (FTPO) method targeting doom loops where models repeatedly output the same span until context exhaustion. The method aims to improve reasoning model reliability.

Apple ML Research proposes Fortress to stabilize search recommendations

The paper addresses temporal instability in recommendation systems caused by volatile input features. Fortress uses temporal data augmentation and feature pruning to improve consistency in multi-stage prediction pipelines.

British Columbia eyes legal action against OpenAI over mass shooting

British Columbia is exploring legal action against OpenAI for failing to alert authorities about threats made on ChatGPT before the February mass shooting in Tumbler Ridge. The case raises questions about AI companies' responsibility to monitor and report dangerous content.

Apple's DynaMiCS fine-tunes LLMs with dynamic domain constraints

Apple ML Research proposes DynaMiCS, a method that uses dynamic data mixtures to fine-tune LLMs while preserving performance on constrained domains like general knowledge and safety. It outperforms fixed heuristics and adaptive baselines.

Ex-Tesla Scientist Unveils Plans For European Humanoid Robot

UMA Robots unveiled a humanoid robot prototype developed in 9 months by a small team. The robot uses a single end-to-end neural network, operates robustly for hours, and has safety baked in at all levels of the full stack.

Psyche Network Architecture enables distributed training without bottlenecks

The Psyche Network Architecture, introduced by Nous Research, claims to enable distributed training without incurring data transfer bottlenecks. Multiple models have already been trained using this method, according to the announcement.

Boston Dynamics' Atlas robot performs at FIFA World Cup 2026

The Atlas humanoid robot delivered the match ball and performed celebrations at the NYNJ Stadium during a Brazil vs. Norway FIFA World Cup 2026 match. Boston Dynamics engineers later detailed the robot's design and the logistics behind the live demonstration.

New method estimates LM memorization, separates from generalization

Proposes a method to estimate how much a language model knows about a datapoint, distinguishing memorization from generalization. Prior studies struggled to disentangle the two.

MT-EditFlow: Reinforcement Learning for Multi-Turn Image Editing

Apple ML Research introduces MT-EditFlow, a reinforcement learning method for multi-turn image editing using flow matching. The approach is designed to handle complex, sequential edits beyond single-turn capabilities.

LawNext podcast: Inside Anthropic's Claude for Legal

Launched in May, Claude for Legal includes over 20 MCP connectors and a dozen practice-area plugins. Anthropic's Mark Pike discusses the product's origins and the debate it reignited.

Apple revisits ASR error correction with compact seq2seq models

Apple's research introduces a compact seq2seq model for ASR error correction with 15x fewer parameters than LLMs, achieving 1.5/3.3% WER on LibriSpeech test-clean/other and outperforming LLMs. The model uses correction-first decoding with ASR acoustic scores and generalizes across CTC, Seq2seq, and Transducer architectures, avoiding latency and hallucination issues common with LLMs.

MiniMax Token Plan Plus offers cost reduction for AI coding tasks

Box survey reveals key divides between enterprise AI leaders and laggards

The survey of 1,640 IT decision makers found that content access, governance, and platform flexibility are the main differentiators. AI leaders prioritize these areas while laggards struggle with foundational data challenges.

Schneider Electric builds LLMOps foundations with LangSmith

Schneider Electric used LangSmith to improve observability, evaluation, and deployment for enterprise AI products. The case study details how they scaled LLMOps foundations across their organization.

Build a serverless image editing agent with Amazon Bedrock AgentCore harness

Tutorial walks through building a serverless image editing agent using Amazon Bedrock AgentCore harness. The harness handles orchestration, tool routing, memory, and compute via configuration.

iOS 27 Beta 3 Adds Pace & Expressivity Sliders for Siri Voices

With iOS 27 developer beta 3, Apple introduces 'Pace' and 'Expressivity' sliders allowing testers to customize Siri's speaking speed and emotional tone. The update gives an early preview of Apple's AI-powered Siri improvements. It's available now to developers enrolled in the beta program.

MindStudio blog series on building reliable AI agents

The series covers 7 components for long-running agents: goals, evaluators, verifiers, loops, orchestration, observability, and memory. Separate articles detail token reduction strategies that cut costs 50-99% and the gate pattern to prevent premature actions.

Hugging Face models now available on Microsoft Foundry managed compute

Hugging Face models can now be deployed on Microsoft Foundry managed compute. The integration enables users to run models at scale with managed infrastructure. Developers can access models directly from the Hugging Face Hub on Foundry.

Brook.ai and SRHO partner to expand AI remote care across US

Brook.ai partners with SRHO, representing 275+ hospitals across 20 states, to deploy an AI-enabled remote care platform. The platform combines clinician-governed workflows with AI to support chronic care management.

Runway launches AI-powered slide generation tool

Invisible data poses hidden risks for AI agents

The article describes 'invisible data'—exceptions, approvals, context, and undocumented institutional knowledge—that can break AI agents in large institutions. This invisible data is more dangerous than bad data, leading to poor agent performance. Organizations need better data strategies to surface these blind spots.

AI and automation reshape healthcare revenue cycle management

AI is boosting productivity in healthcare revenue cycle management (RCM) by 10-20%, with another 20% improvement expected within two years, per Omega Healthcare CEO Anurag Mehta. Voice AI agents from SuperDial are also automating billing calls between providers and payers, handling large volumes.

NVIDIA discusses quantum computing infrastructure at ISC

ABB Robotics launches Flexley Stack F712 autonomous forklift with vSLAM

The Flexley Stack F712 uses vSLAM navigation for storage and retrieval operations. It is the latest addition to ABB's mobile robot portfolio, creating a more complete ecosystem.

Treasury internal report warns of AI bubble dangers

A Treasury Department internal report flags risks of an AI investment bubble, warning that overvaluation could lead to economic instability. The report emphasizes caution amid soaring AI spending.

Discord admits AI moderation bug wrongfully banned 8,000 users

A bug in Discord's AI moderation system flagged harmless images like spreadsheets and transparent backgrounds as harmful, causing over 8,000 wrongful bans in two months. Discord acknowledged the issue and is working on a fix.

The real cost, security, and culture problems behind enterprise AI agents

At VentureBeat's AI Impact event, Red Hat senior director Brian Gracely outlined key obstacles enterprises face when scaling agentic AI, including hidden costs, security liabilities, and cultural resistance. The discussion focused on what separates successful deployments from stalled pilots.