6 min read · 15 sources

Google released Gemini 4 Carbon, targeting vulnerability discovery, exploit verification, and long-horizon software engineering. Rather than chasing generic chat benchmarks, Google pointed the model at real-world infosec workflows, where it scored 88.2% on SEC-bench Pro, 54.8% on ExploitGym, and 71.0% on CWE-bench. It also took the top slot on the freshly minted Artificial Analysis Intelligence Index v5.0 with an aggregate score of 59.

The pricing structure is built for agent loops that churn millions of tokens debugging across multi-day lifecycles. Standard pricing sits at $2 per million input tokens and $10 per million output tokens, but prompt caching knocks 95% off the input bill. For teams operating security fuzzers or autonomous code remediation pipelines, that brings warm-context iteration down to $0.10 per million input tokens.

Repurposing Gated DeltaNet layers as exact RAM turns the forward pass of an open-weight LLM into a deterministic RISC-V computer without an external interpreter.

Running RISC-V Inside Qwen 3.5 Attention Layers

Source: divagr.com ↗

A developer built Qwenputer, which embeds an entire, deterministic RISC-V processor directly into the weight tensors of an unmodified Qwen 3.5 4B model. The execution requires zero external interpreters, tool calls, or Python runtime environments; the compute happens entirely inside the model’s native tensor math.

The trick exploits the linear recurrence properties of the architecture’s Gated DeltaNet (GDN) layers. By fixing the decay gate and write beta values to exactly 1 and routing state through one-hot key addressing, the developer converted the attention mechanism into an addressable, persistent read/write RAM. The model evaluates arbitrary arithmetic, runs compiled C, and executes MicroPython across successive token generations. For security researchers, this is a clean proof-of-concept for running sandboxed logic inside model weights - or hiding reasoning steps from standard chain-of-thought monitoring.

NVIDIA Ships Nemotron 3.5 Transcribe and Granary v2

Source: huggingface.co ↗

NVIDIA launched Nemotron 3.5 Transcribe, releasing an open-weight model alongside its Granary v2 curation pipeline. The model takes the top spot on the Hugging Face Open ASR Leaderboard for English, transcribing 40 languages and handling two-way translation for 31 languages against English.

Crucially for audio infra teams handling raw media streams, the checkpoint folds speaker diarization directly into the core forward pass. It tracks up to eight distinct speakers across 13 languages on inputs up to one hour long without needing a secondary clustering pipeline like PyAnnote. The release includes fine-tuning recipes via Granary v2; NVIDIA showed that target domain fine-tuning drove a 45.8% relative Word Error Rate (WER) reduction on specialized medical speech.

The Benchmark Dilemma: Artificial Analysis Index v5.0

Source: interconnects.ai ↗

Artificial Analysis rolled out version 5 of its Intelligence Index, attempting to tame the broken state of frontier LLM evaluation. The v5 specification splits weight evenly across four quadrants: Knowledge Work (25%), Coding (25%), Science (25%), and General Capabilities (25%). It swaps out saturated multiple-choice tests in favor of rigorous execution benches like FrontierCode 2.0 and Terminal-Bench 4.0.

The release underlines the tension every infra team faces when sizing models for production: composite metrics mask domain-specific collapses. While Gemini 4 Carbon took the #1 rank, single-number indexes obscure the trade-offs between context window stability, tool-calling determinism, and latency variance. When billions in compute spend hinge on fractional index movements, benchmark gaming remains an operational risk for anyone buying API access based purely on leaderboard position.

Autonomous Inference Stack Optimization

Source: x.com ↗

Serving high-throughput models usually means throwing systems engineers at vLLM configurations, kernel patches, and speculative decoding draft models. Decagon outlined AutoResearch, an autonomous agent system designed to run continuous tuning across its inference fleets.

The setup pairs Devin Cloud for agent orchestration with Modal for spinning up elastic GPU instances. The agent identifies production serving bottlenecks, reconfigures memory layouts, benchmarks speculative draft strategies, and verifies output latency by replaying sanitised production traffic traces. Production changes still require a human sign-off, but automated exploration of compiler flags and batching parameters delivers up to 2.5x throughput gains within fixed latency SLAs, maximizing requests per second per GPU.

Vercel CI Launches to Handle Agent PR Floods

Source: vercel.com ↗

Autonomous coding agents write code faster than traditional Continuous Integration can process it, creating massive backlogs of pending runs. To address this, Vercel announced the beta of Vercel CI, a task-level CI engine built to handle high-frequency commits without ballooning compute bills.

Instead of writing brittle, verbose YAML workflow files that fall out of sync with your monorepo, Vercel CI reads existing Turborepo configurations. Appending "tags": ["ci"] to tasks in turbo.json enables the runner to skip unaffected packages automatically using Remote Cache. Jobs execute on Vercel’s isolated Hive compute fleet, shrinking wall-clock times from git push to green checks by 2x to 6x.

Model Routing: Microsoft Copilot Harness and GitHub Copilot Fusion

Source: blogs.microsoft.com ↗

Microsoft and GitHub both shipped infrastructure updates focused on dynamic model orchestration to stop developers from defaulting to the most expensive model for every prompt.

Microsoft opened the Copilot harness via Microsoft Foundry. Originally written internally by GitHub to back Copilot Workspace and the CLI, it exposes an enterprise-grade runtime for routing multi-turn agent logic across heterogeneous models while preserving state through the Model Context Protocol (MCP).

Simultaneously, GitHub launched Copilot Fusion into General Availability. Fusion provides three dynamic routing tiers: Standard, High, and Max. The engine uses model cascades and reflection passes; Fusion Max hits a score of 60 on the Artificial Analysis Coding Agent Index, rivaling Claude Code running Opus 5.5 Max while slashing token consumption by 50% and executing runs 40% faster.

Alongside Fusion, GitHub also enabled self-improving loops for Copilot Automations. Meta-agents observe merged PRs, rejected suggestions, and code review comments, continuously proposing pull requests to tighten the underlying automation prompts and tool definitions.

Enterprise AI Spend and Attribution

Source: techcrunch.com ↗

Anthropic rolled out new spending and attribution tools aimed at finance teams trying to map API burn to engineering velocity. The tooling arrives alongside an eye-popping data point: an external survey of 276,000 developers by Jellyfish found that Claude Code now authors roughly one-third of all committed code revisions among teams using the platform.

Anthropic reported 900% net revenue retention, with developer tooling spend climbing 10x year-to-date to claim roughly 2% of total engineering budgets at adopting firms. The new dashboard gives managers granular token tracking down to the project and failure level, preventing runaway loops from chewing through shared API keys.

Quick Links: Audio, SMS Interfaces, and Autonomous Labs

Source: elevenlabs.io ↗

• ElevenLabs Cue: ElevenLabs open-sourced architectural details on Cue, its character-level forced alignment engine. The model handles an hour of speech in under a second across 29 languages, aligning raw waveforms to transcript characters with millisecond accuracy for real-time subtitling and voice synchronization.
• OpenAI Texts You: OpenAI launched a preview letting US ChatGPT Pro users message their Dot agent over SMS, iMessage, and RCS. Built with Linq, the interface unifies context with web chat sessions and allows the agent to initiate outbound alerts.
• Google Connected Apps: Google introduced the Connected Apps Platform, letting teams expose their services into Search and Gemini via cloud or on-device MCP servers mapped through AppFunctions APIs.
• Lor Research Agents: Mirendil opened the private beta for Lor, an autonomous agent platform that spins up parallel experiments across hundreds of nodes, writing ML code, running training runs, and parsing evaluation logs for days without human intervention.
• Meta Waddle World: Meta launched Waddle World, a 16-bit virtual pet app in the Philippines. The Tamagotchi-like journal deliberately avoids productivity features or open-ended romantic companion dynamics, testing low-liability consumer interaction loops.