Search the archive

Try IronClaw Free [13]→

ANTHROPIC LAUNCHES ADVISOR TOOL FOR CLAUDE PLATFORM API USERS

The new advisor tool on Anthropic's Claude Platform API provides developers with the ability to use Opus as an advisor alongside Sonnet or Haiku as executors. This allows agents to access advanced reasoning capabilities while maintaining operational costs at the level of the more efficient executor models. The advisor tool is now publicly available through a configuration in the Messages API request.

MULTIMODAL EMBEDDING & RERANKER MODELS WITH SENTENCE TRANSFORMERS

The v5.4 update of Sentence Transformers introduces multimodal embedding and reranker models that enable encoding and comparing texts, images, audio, and videos within a shared embedding space for tasks like cross-modal search and retrieval-augmented generation. These advanced models map different modalities into a consistent embedding space and assess relevance across mixed-modality pairs, though lower cross-modal similarity scores may occur due to modality clustering. Notably, multimodal rerankers provide high-quality scoring by evaluating individual pairs for more accurate retrieval results.

PROCESS DRIVEN IMAGE GENERATION

Meta has introduced a multi-step framework for image synthesis that alternates between textual reasoning and visual updates. This enables models to iteratively plan, draft, reflect, and refine outputs.

EFFICIENT ROLLOUT SCALING FOR DIFFUSION RL

NVIDIA's Sol-RL introduces a two-stage framework that separates exploration and training, using FP4 rollouts to generate large candidate sets and BF16 for selective policy updates. This approach reduced compute costs while improving alignment and accelerating convergence in diffusion model post-training.

DARIO SAYS CONTINUAL LEARNING IS SOLVED. IS IT?

Continual learning research is aimed at breaking past the feasible horizon of current techniques. If models can't learn new things while performing tasks, they will struggle when the task horizon grows very long. Context management and other engineering improvements can push task horizons to weeks, or even months, but very long-horizon tasks remain out of reach. Human-level continual learning may be 'solved', but the revolution is not yet complete.

INTRODUCING KELLYBENCH

KellyBench evaluates sequential decision-making in sports betting by simulating the 2023-24 English Premier League season. Models like Claude Opus 4.6 and GPT-5.4 struggled, with none achieving positive returns, highlighting their limitations in adapting to long-term strategies. This underscores a need for complex environments enabling agents to learn from experience under uncertainty.

FULLY REMOTE)

TLDR's Applied AI team is tasked with making every process at TLDR legible to code, runnable by anyone, and composable into larger workflows. Join a small, fast moving team using the latest AI tools with an unlimited token budget. Learn more [21].

THE GEMINI APP CAN NOW GENERATE INTERACTIVE SIMULATIONS AND MODELS

The Gemini app now supports generating interactive visualizations, like 3D models and simulations, directly in chat.

GOOGLE'S PAPERORCHESTRA AI CONVERTS LAB NOTES INTO PUBLICATION-READY RESEARCH PAPERS

Google Cloud AI's PaperOrchestra converts disorganized lab notes into submission-ready academic papers using five specialized agents.

OPENAI PROJECTS STEEP ADVERTISING GROWTH, TARGETING $100 BILLION BY 2030

OpenAI expects its advertising business to generate $2.5 billion in revenue this year.

Try Wispr Flow Free [4] | Get Flow

META INTRODUCED MUSE SPARK

Meta introduced Muse Spark, a multimodal reasoning model with tool use, visual chain-of-thought, and multi-agent orchestration, as part of its push toward personal superintelligence.

SCALING MANAGED AGENTS: DECOUPLING THE BRAIN FROM THE HANDS

Harnesses encode assumptions about what Claude can't do on its own. These assumptions need to be frequently questioned as models improve. Anthropic built Managed Agents as a system for 'programs as yet unthought of'. It is designed to accommodate future harnesses, sandboxes, or other components around Claude.

ANTHROPIC MANAGED AGENTS

Anthropic's Managed Agents is a hosted system that separates agent interfaces from underlying implementations to support long-running tasks as models evolve.

INTRODUCING LEARN MODE: YOUR PERSONAL CODING TUTOR IN GOOGLE COLAB

Google Colab has enhanced its Gemini integration with Custom Instructions and Learn Mode, enabling personalized AI assistance and coding guidance. Custom Instructions allow users to adjust Gemini's behavior to fit their workflow or project needs, and Learn Mode provides step-by-step coding support instead of complete solutions. These updates enhance user control and facilitate skill-building, sharing personalized AI settings with the Colab community.

INSIDE THE AI INDUSTRY'S MOST EXPENSIVE MISTAKE

An internal leaderboard at Meta that ranked employees on token usage revealed that, over a 30-day period, the company used around 60 trillion tokens. It is estimated that all books published amount to about 20 trillion tokens. The AI industry seems obsessed with token spend, but the metric is easily hacked, and more token use doesn't mean more value is provided.

SYSTEMS ENGINEERING: BUILDING AGENTIC SOFTWARE THAT WORKS

Building reliable agentic software requires treating it as a full system, not optimizing isolated components like tools or storage. This post outlines five critical layers that must be designed together to avoid cascading constraints and failures. Using a real open-source example, it shows how structured data, enforced permissions, and consistent interfaces enable agents to improve over time and operate safely in production.

MONARCH: AN API TO YOUR SUPERCOMPUTER

Monarch is a distributed programming framework for PyTorch that makes running distributed training jobs on huge clusters easy. It makes clusters programmable through a simple Python API that exposes the supercomputer as a coherent, directly controllable system. Monarch is optimized for agentic usage, and it provides consistent infrastructure abstractions and exposed telemetry. It can turn a dev machine into a supercomputer, leveling up its agents.

CLAW-EVAL BENCHMARK FOR AI AGENTS

Claw-Eval provides a human-verified benchmark for evaluating LLM agents across 139 real-world tasks using Docker sandboxes, multiple services, and structured grading.

BUGBOT NOW SELF-IMPROVES WITH LEARNED RULES

Bugbot's resolution rate now nears 80%, significantly outperforming other AI code review products. Bugbot harnesses real-time signals from past runs to self-improve by transforming feedback into learned rules. Over 110,000 repositories use this feature to generate over 44,000 learned rules, enhancing Bugbot's focus on specific issues and business context.

ANTHROPIC LOSES APPEALS COURT BID TO TEMPORARILY BLOCK PENTAGON BLACKLISTING

A federal appeals court has blocked Anthropic's request to temporarily block the Department of War's blacklisting of the company as a lawsuit challenging the sanction plays out. Anthropic is now excluded from the Department of War's contracts, but it can continue working with other government agencies while the litigation plays out. Defense contractors will be prohibited from using Claude in their work with the agency, but they can use it for other cases.

AI AGENT POKE MAKES SETTING UP AUTOMATIONS AS EASY AS SENDING A TEXT

Poke, a new AI agent accessible via text messaging apps, simplifies automation with pre-made "recipes" for tasks like scheduling and smart home control. Backed by $25 million in funding and valued at $300 million, Poke aims for widespread adoption by offering flexibility in pricing and encouraging user-generated automation recipes. Unlike complex systems like OpenClaw, Poke's user-friendly approach targets a broader audience without compromising on functionality.

credits*. REGISTER NOW!

APPLE SHOWS ITS CARDS, PLANS TO MOVE THE PRODUCTION OF ITS UPCOMING BALTRA ASIC IN-HOUSE

Apple has a well-established proclivity for vertical integration.

ALTK‑EVOLVE: ON‑THE‑JOB LEARNING FOR AI AGENTS

ALTK-Evolve enhances AI agents by transforming raw agent trajectories into reusable guidelines, improving reliability in complex tasks without increasing context bloating.

PERPLEXITY'S SHIFT TO AI AGENTS BOOSTS REVENUE 50%

Perplexity's move from AI-powered search to AI agents has reportedly paid off.

THIS WILL BECOME THE DEFAULT WAY TEAMS USE NOTION + CLAUDE

Run roadmaps and task boards inside of Notion, assign out tasks to Claude Agents, and then collaborate with the agents and team to review the work from inside of Notion all the way through to merged PRs in GitHub.

ENTERPRISE AI ADOPTION

OpenAI reported rapid enterprise growth, with AI moving beyond experimentation into core business workflows and a strategy centered on unified agents and a company-wide AI layer.

Get the guide to Advancing physical AI.

PROJECT GLASSWING: SECURING CRITICAL SOFTWARE FOR THE AI ERA

Anthropic's Claude Mythos Preview autonomously identified thousands of zero-day vulnerabilities across major operating systems and browsers. Project Glasswing, in partnership with major tech companies, uses these capabilities to enhance cybersecurity by detecting and fixing vulnerabilities at scale. Anthropic plans to develop safeguards and broaden industry cooperation to address security challenges in the AI era.

GLM-5.1: TOWARDS LONG-HORIZON TASKS

GLM-5.1 is a flagship model for agentic engineering created by Z.ai. It achieves state-of-the-art performance on SWE-Bench Pro. The model is built to stay effective on agent tasks over much longer horizons than previous generations. It can sustain optimization over hundreds of rounds and thousands of tool calls. The model breaks complex problems down, runs experiments, reads results, and identifies blockers with real precision.

MY PICTURE OF THE PRESENT IN AI

Ryan Greenblatt is the chief scientist at Redwood Research, a research organization with the mission of aligning superhuman AI. This post goes through some of his best guesses for the current situation of AI. The scenario forecast discusses R&D access regulations, engineering capabilities and qualitative abilities, misalignment and misalignment-related properties, cyber, bioweapons, and economic effects. Some of the claims are highly speculative, while others are better grounded.

CLAUDE MYTHOS

Anthropic detailed early evaluations of Claude Mythos Preview, highlighting strong performance in discovering zero-day vulnerabilities and reverse-engineering exploits, prompting a coordinated security initiative called Project Glasswing.

AI CAN'T READ AN INVESTOR DECK

Current AI models struggle with interpreting complex financial documents, especially with visual data extraction. GPT-5.4, Gemini 3.1 Pro, and Claude Opus 4.6 consistently falter when processing dense charts and images, only achieving 56% to 64% accuracy compared to 72% to 80% with text-only inputs. These findings highlight significant gaps in AI's ability to perform real-world financial reasoning tasks, making the displacement of financial analysts with AI appear premature.

NOW!

FASTER MOE INFERENCE WITH WARP DECODE

Cursor's “warp decode” is a kernel design that reorganizes MoE inference around output neurons instead of experts. It achieves ~1.8x higher throughput and improved numerical accuracy on Blackwell GPUs.

TORCHTPU: RUNNING PYTORCH NATIVELY ON TPUS AT GOOGLE SCALE

Google's Tensor Processing Units (TPUs) are foundational to the company's supercomputing infrastructure. The company's custom ASICs power training and serving for Google and its Cloud customers. TorchTPU is a stack that makes it easy for the AI community to access the full capabilities of TPUs. It provides the APIs and tools needed to extract every ounce of compute from Google's hardware. This post takes a look under the hood at the engineering principles behind TorchTPU.

TRIATTENTION FOR KV CACHE COMPRESSION

TriAttention estimates KV importance in pre-RoPE space using stable Q/K centers and distance-based scoring, preserving long-context reasoning quality while sharply reducing KV memory use and improving throughput.

META AI SCALES RL FOR ML ENGINEERING AGENTS

SandMLE is a framework for building small but structurally realistic MLE environments that made on-policy RL practical for ML engineering agents by cutting execution cost more than 13x.

WE'RE ACTUALLY RUNNING OUT OF BENCHMARKS TO UPPER BOUND AI CAPABILITIES

METR's Time Horizon suite is being saturated. Frontier AI models can reliably do all but maybe a dozen or so tasks in the suite, making it hard to upper bound their time horizon. New benchmarks are becoming more expensive to grade and create. The situation will likely get worse as AI progress continues. It is likely that, by mid-2027, no benchmark score from a 2026 or earlier benchmark will be able to rule out dangerous capabilities from frontier AI systems.

ELON MUSK ASKS FOR OPENAI'S NONPROFIT TO GET ANY DAMAGES FROM HIS LAWSUIT

Elon Musk's lawsuit against OpenAI is expected to go to trial later this month in Oakland, California. Musk has amended the lawsuit to ask that any damages he might win be awarded to OpenAI's charitable arm rather than to himself. The amendment also asks that OpenAI CEO Sam Altman be removed from the OpenAI nonprofit's board. Musk is seeking more than $150 billion in damages from both OpenAI and Microsoft as he believes OpenAI strayed from its non-profit mission and defrauded him as a donor in seeking to convert to a for-profit company.

FULLY REMOTE)

TLDR's Applied AI team is tasked with making every process at TLDR legible to code, runnable by anyone, and composable into larger workflows. Join a small, fast moving team using the latest AI tools with an unlimited token budget. Learn more [18].

GOOGLE CONTROLS THE MOST AI COMPUTING POWER, DRIVEN BY ITS CUSTOM TPUS

Google holds around 25% of all compute sold since 2022.