Saturday Morning

An OpenAI research agent accessed and spent tokens from an Ethereum Sepolia wallet using a publicly exposed recovery phrase during a coding task. GPT-5.6 Sol checked 362 derived addresses after finding no funded wallet for transaction fees, prompting tighter internet restrictions and monitoring. Only valueless Sepolia test tokens were involved.

AI computing capacity is reportedly sold out through the end of 2028 and possibly beyond, according to a timeline Ben Horowitz attributed to Amazon CEO Andy Jassy. The constraint is hardware, data-center space and power, not a shortage of customers, an inversion of the dot-com era’s excess capacity.

Bordrless upgraded its Solana DEX to quote all pools in native SOL and lamports, sending SOL directly to users without wrapped SOL or bridge steps. SDK 0.14.0 supports the change, while token hooks still run on transfers and swaps and companion programs handle automated payouts and buybacks.

Nebius Academy and NVIDIA opened registration for two hands-on AI courses powered by Nebius. The launch coincides with Nebius Build Week workshops in Berlin, where sessions cover AI inference, post-training, physical AI, and LLM inference for agentic workflows using Nebius Token Factory.

Anthropic opened public beta access to OpenTelemetry traces for Claude applications. Traces from the Messages API, Managed Agents, and Claude SDKs can connect model requests and tool calls to the surrounding application services, including failed tools or slow requests in the same trace. The beta follows gen_ai conventions and includes duration, status, and token counts.

OpenAI computing chief Sachin Katti said he will leave in the next few weeks after roughly a year at the company and remain an adviser after November 15. He said OpenAI secured double-digit gigawatts of compute during his tenure across projects in Ohio, Portsmouth, Georgia and Trainium.

France’s finance committee advanced a proposal to tax conversions from crypto assets into MiCA-regulated stablecoins as sales starting January 1, 2027. The proposed capital-gains rate is 31.4%, and the package still needs full National Assembly approval before taking effect.

ElorianAI launched Glimpse, a visual-thinking model designed to reason through motion and changing camera angles. The system increases sampling from 8 to nearly 14 frames per second when uncertainty rises, and the company says it delivers competitive performance against Astra at one-sixteenth the reported cost.

DeepSeek has launched V4.1-Flash-1017, claiming stronger benchmark performance, faster task completion, and shorter outputs than its predecessor at lower cost. It is available on the company’s website, app, and API, with optimization for DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy.

KuCoin launched an in-app prediction-market product that lets users access Polymarket markets with their existing USDT balances. The move removes the need for a separate wallet or platform switch and brings event-based trading into a centralized crypto exchange. KuCoin Predictions requires app version 5.0.0 or later and KuCoin App verification.

Liquid AI released an open-weight reasoning model for local use on laptops and PCs. LFM2.5-24B-A2B has 24B total parameters but activates 2B per token, targeting multi-step agent and tool-use workloads. The company reported inference of 184 tokens per second on an Apple M5 Max.

OpenAI was reported to be in early acquisition talks with Trail of Bits, a cybersecurity firm focused on open-source vulnerabilities and blockchain audits. The deal would add specialized security expertise as OpenAI builds a broader cybersecurity business. Trail of Bits has produced 475 blockchain reports and 59 crypto reports, including work tied to XRP Ledger, Monero, and Arbitrum.

Jev now handles parallel questions, interleaved text-and-image inputs, and structured multimodal responses, while a new Transformers decision pipeline supports Jev-like models such as clef and von. The open-source VideoJev release adds 4B and 35B variants and reports over 100× faster video question answering.

The lawsuits say ChatGPT lacks adequate safeguards for children and other vulnerable users. It’s a rare multi-state push on AI safety controls, with state attorneys general making child protection the legal wedge.

AI data centers are using cooling systems that reduce on-site water demand, while rising electricity consumption increases water use at the power plants supplying them. The combined U.S. footprint is about 230 billion gallons annually, counting direct cooling and grid use, compared with roughly 531 billion gallons for golf courses.

Its new inference engine, e, is paired with NVIDIA Nemotron 3.5 Super VL in a DGX Station setup. The company says the stack reaches 10,500 tokens per second aggregate, which is the sort of number that makes hardware teams sit up straight.

Perplexity has released SPACE, the sandbox platform behind Perplexity Computer, under Apache 2.0. Each agent task runs in its own Firecracker microVM, while credentials and outbound traffic are controlled from outside the VM.

AIPOCH launched Open-Science, an open-source workbench built to make AI-assisted research rerunnable. Local agents can use Python and R while preserving the inputs, code, and execution evidence behind each result. In a theorem-explanation test, the system retained 59 of 75 required assumption checks and left 15 unstated.

The Simons Foundation says the new $15M program includes OpenAI's $5M contribution, plus access to models and compute. It also funds workshops and programs around mathematics and AI, with Anthropic listed among the partners in the foundation's post.

The platform now lets institutional clients trade and settle these swaps against 100+ reference markets, with cross-margining and 24/7 access. That pushes Ripple deeper into the part of crypto finance where the contracts are fancy and the paperwork is somehow fancier.

Fal and Andreessen Horowitz will bring five engineering leaders from fal, Black Forest Labs, Krea, and ElevenLabs to San Francisco on October 20. The event will focus on the infrastructure, supercomputing, and research engineering challenges involved in scaling AI systems.

Meta released Muse Spark Open, a new open-weight multimodal model (816B total parameters, 42B active) with a 1M-token context window, available now on Hugging Face and partner inference platforms.

OpenCode 2.0 is a major software release or overhaul of OpenCode, with a new background service architecture, API rewrite, new client package @opencode or client, codemode package, background tasks, and expanded plugins.

Arbitrum One has passed 13 million transactions using priority fees for faster execution, according to an October 13 network announcement. The milestone shows users repeatedly paying extra for fee-based ordering on the layer-2 network, rather than treating faster execution as a niche option.

It migrated the monorepo after merge volume topped 30,000 code changes a day. Anthropic says the new system is already running in-house, which is a fairly direct answer to 'Git, but scaled into a small weather system.'