On Snapshot, the research proposal asking whether Starknet should continue its transition into an independent L1 is currently passing with more than 345 million votes in favor (99.95%), against roughly 157,700 votes opposed (0.05%). Voting is scheduled to close on October 22.

The tally is a near-unanimous signal from participating voters on a structural architectural question — whether one of the largest ZK rollups keeps moving toward standalone L1 status — although it is a research proposal and does not itself execute any upgrade.

Snapshot is an off-chain, gas-free voting tool, and the outcome of this research proposal is a sentiment signal rather than an on-chain change. StarkWare has said the L1 move remains under evaluation rather than decided, describing it as the network's clearest strategic path forward.

On October 18, 2026, Anthropic said it plans to commit 5% of its shares to the newly established Anthropic Public Benefit Fund, an entity structured so that its stake increases as the company grows. The commitment is subject to stockholder approval.

Anthropic said the fund will have greater flexibility and scale, allowing it to explore options such as large donations to traditional nonprofits or direct distributions to specific populations.

The fund's stake is designed to increase as the company's valuation grows, rather than remaining fixed at 5%. The arrangement has not yet been approved by stockholders.

After chatbot startup Character.AI filed for bankruptcy earlier this week, Disney agreed to hire the majority of its workforce and to acquire some of the company's assets for $20 million. Character.AI's former CEO, Karandeep Anand, had already joined Disney as its first chief technology officer.

The arrangement shifts most of a well-known consumer AI startup's team into a major entertainment company, following Disney's earlier hire of the startup's chief executive into a newly created CTO role.

The $20 million figure covers some of Character.AI's assets, according to the startup's earlier statement. The report references a legal matter involving Character.AI, but its details were not included.

Anthropic has been pitching U.S. lawmakers on an unpublished draft bill that would let the government seek restrictions on a frontier model's release or use only when the model poses a foreseeable and imminent risk of more than $1 billion in damages or at least 50 deaths or serious injuries, according to POLITICO. The draft would also require regular independent evaluations of frontier models and would place some state-level AI safety rules under a federal framework.

The draft sets explicit, high thresholds for when Washington could intervene on a model, and its preemption element would move part of state-level AI safety oversight into a federal framework. It is an unpublished proposal being pitched to lawmakers, not introduced or enacted legislation.

The risks cited as triggers include biological weapons, loss of developer control and cyberattacks. The text remains an unpublished draft circulated to lawmakers rather than a filed bill.

Crypto.com and Novig filed a lawsuit against Missouri over the state's treatment of sports event contracts. Kalshi, Polymarket and Robinhood held off on their own moves amid the resulting regulatory uncertainty.

The suit puts the question of whether states or the federal Commodity Futures Trading Commission governs event contracts traded on registered exchanges before the courts, while major platforms' hesitation shows how that unresolved question is already shaping product rollouts.

This is a filed lawsuit rather than a ruling, and the report gives no court, judge or specific legal claims. The CFTC holds what the Congressional Research Service describes as exclusive jurisdiction over futures, options and swaps traded on registered designated contract markets, and how far that preempts state regulation of event contracts is being litigated.

VAST has open-sourced EdenLive, a real-time interactive world model that the company says runs at 720P resolution and above 24 FPS on a single consumer-grade GPU. The release is positioned around letting users freely create and experience interactive worlds.

Running an interactive world model at 720P and above 24 FPS on one consumer-grade GPU lowers the hardware threshold for developers and creators who want to run such models locally rather than on server clusters.

The 720P and 24+ FPS figures come from VAST's own release claims; independent benchmark results were not part of the announcement. The model is released as open source.

Bloomberg reported on October 18 that CoreWeave's high-yield bond yields rose to 15% this week, a level typically associated with debt distress, as high-yield investors grow increasingly reluctant to fund technology startups. The concerns have spread to the broader high-yield market, where the average yield climbed to 8.41% on Wednesday, an 18-month high, and the risk premium widened to its highest level since March.

Shorecliff Asset Management chief investment officer Grant Nachman said that at current market levels, unsecured financing for large "neoclouds" may be too expensive to complete, or may not be completable at all, and that new deals may require backing from assets or contracted cash flows plus stronger creditor protections.

CoreWeave says its financial position is "fully sufficient" to cover its debt obligations, and Nvidia CEO Jensen Huang has publicly defended the company's financial health. Some investors argue the core issue is not a sudden deterioration in corporate credit quality but the large volume of new debt supply still expected over the next 18 months.

The Cysic Foundation said its cross-chain bridge was exploited on October 17, when an attacker gained unauthorized access to a database supporting the bridge service and used a processing logic flaw to submit a fake mint message, creating 10 million CYS on BNB Smart Chain. The unauthorized mint occurred at 12:31:38 UTC, and Cysic added the recipient address to the token contract's blacklist at 13:47:52 UTC.

About 2.8 million CYS remained at the blacklisted address and can no longer be transferred out. Cysic said the incident was confined to bridge message processing and minting, and did not affect the mainnet, other operational systems, or existing assets in user wallets.

Cysic said the 10 million tokens minted does not by itself establish the final dollar loss, and that transfers, swaps and actual losses remain under review; roughly 160.8 million CYS were in circulation at the time, according to CoinMarketCap data cited in reports. The foundation has hired an unnamed security firm to investigate the attack path and trace the funds, and said the bridge will reopen only after a security review is completed.

Binance is shutting down its Bitcoin mining pool after the pool's hashrate dropped roughly 65%, ending the exchange's mining-pool service for customers.

Miners who directed hashpower to Binance Pool lose that option and must route their machines to other pools, removing one well-known exchange-operated pool from the choices available to them.

The roughly 65% decline refers to the Binance pool's own hashrate, not Bitcoin's network-wide hashrate. The pool was marketed as an exchange-integrated service with instant reward access, VIP perks and cloud mining features.

Together Computer released Hybrid Reasoning Agents (HybridAgents), an agent architecture in which the model performs its reasoning in continuous latent space and emits text only for actions and tool calls. The release ships with a GitHub repository, a HuggingFace model collection, and evaluations in arXiv paper 2610.18861.

The authors position the approach as bridging latent reasoning, which is efficient internally but lacks native tool access and reasoning transparency, with text-based agents that can act and be inspected. It is an architectural departure from explicit chain-of-thought rather than a new model checkpoint alone.

The open release spans code on GitHub (togethercomputer/HybridAgent), a HuggingFace collection, and a published evaluation section in the arXiv HTML version. In the discussion thread, one user reported that the 14B variant, run through llama.cpp with bartowski's 8-bit quantization, looped on variations of the same paragraph for roughly five minutes until it exhausted its turn budget on a simple chat, web search and summarization request; the user noted llama.cpp may need updates to support the model.