Vitalik said in a post on X that a new kind of high-level programming language worth trying would compile to Lean, HOL, or similar systems, with the main goal of making definitions and theorems easier for humans to read. He argued that the proof itself matters less from a readability standpoint because it only needs to be correct, while the core value lies in understanding the definitions and the theorems being stated. In the use case he described, AI could generate a long proof, and readers would still need a clear way to see exactly what claims were proved. The comment centers on readability and interpretation rather than the mechanics of proof generation.

Google said Gemini 3.5 Pro is now being tested with partners and will be rolled out more broadly as soon as it is ready. The company also said it has started pre-training Gemini 4, signaling that work on the next generation of the Gemini model family is already underway. Alongside that update, Google announced Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber. Pricing details were disclosed for two of the new offerings. Gemini 3.5 Flash-Lite is priced at $0.3 per million input tokens and $2.5 per million output tokens, while Gemini 3.6 Flash is priced at $1.50 per million input tokens and $7.50 per million output tokens. The update was carried by Odaily, citing Jin10.

Nvidia said its major customers have begun testing Vera Rubin equipment, according to a Bloomberg report cited by BlockBeats on July 21. The company also said its new Vera processor is faster than AMD's Turin. Nvidia added that the chips are being delivered on schedule and are intended for use in artificial intelligence data centers. The update points to progress in the rollout of Nvidia's next-generation hardware for AI infrastructure, based on the company's own statements and the Bloomberg report referenced in the original brief.

ChangXin Tech’s final strategic placement subscription results are out, showing a sharp drop from the initial allotment size. The 30 strategic investors were originally allocated 3.344 billion shares, but the final number actually subscribed came in at 1.667 billion shares, roughly half the initial figure. The change has prompted market speculation, though the report said the outcome was tied to a common industry practice rather than a sudden shift in commitment. According to the report, strategic placement structures are often designed around the upper regulatory limit at the outset, leaving room for later adjustment. For a deal the size of ChangXin Tech’s issuance, the ceiling for strategic placement was 3.344 billion shares, or 50% of the total issuance, with shares later subject to reallocation based on actual conditions. Even though the number of shares ultimately subscribed declined, the 30 strategic investors’ actual subscription amount remained unchanged from their previously committed amount at 14.437 billion yuan. The gap in shares was moved to the online subscription tranche, which significantly raised the winning rate for retail investors participating in the online offering. The item was reported by Jin10 and carried by Odaily.

Pump.fun has introduced a new BOOST mode and will make it the default launch setting for all newly issued Pump.fun tokens going forward. The feature is designed to address the long-running problem of “dead liquidity” during token migrations by channeling permanently locked liquidity back into the market through an automatic buyback-and-burn process. According to Pump.fun, more than $100 million in liquidity has been permanently lost each year during migrations, leaving capital unable to support ongoing market liquidity. Historically, around 20% of liquidity for each migrated token has remained trapped in liquidity pools even after traders fully exited their positions. Under BOOST, those stranded funds will be redeployed within five minutes after migration. The system will carry out buybacks using a time-weighted average price, or TWAP, and automatically burn the purchased tokens. Pump.fun said BOOST will inject 17.6 SOL for SOL pairs and $2,516 for USDC pairs. The feature does not require manual activation and will automatically apply to all new Pump.fun tokens that complete migration after 10:23 a.m. Eastern Time on July 21. Tokens migrated earlier, as well as tokens launched through Mayhem, are excluded.

Decentralized AI platform Venice said its API is being adopted by a growing number of AI applications, according to a Techub News update. Erik Voorhees, the platform’s founder, said on social media that developers looking to build apps without compromising user privacy can try building with the Venice API. Venice positions itself as a privacy-focused decentralized AI platform founded by longtime crypto industry figure Erik Voorhees. The platform says it offers permissionless, censorship-resistant AI inference services and lets users access open-source AI models while keeping data privacy intact. It also says user interaction data is not collected or stored by centralized institutions.

US Treasury prices fell as yields on 10-year and 30-year government bonds climbed to roughly two-month highs, according to Odaily. On Tuesday, yields across maturities were up by 2 to 4 basis points, with the 10-year yield briefly reaching 4.64%, its highest level since late May. The move came as a sharp rise in crude oil prices added to concern that inflation pressure could build and push the Federal Reserve toward rate hikes. At the same time, the US dollar strengthened against most major currencies as the United States and Iran exchanged strikes for a 10th straight day. The combination of higher yields, firmer oil prices and a stronger dollar reflected a broader shift in market positioning around inflation risk and rate expectations.

Changxin Technology’s preliminary offline placement results show that 113 private fund managers, through 2,459 products, received allocations totaling 161 million shares worth 1.436 billion yuan. The filing divided institutional investors into Class A, which included public funds, social security funds, pensions, annuities, bank wealth management, insurers and QFII, and Class B, which included private funds, broker proprietary desks, trusts and finance companies. Class A investors received 1.978 billion shares, accounting for 91% of the offline issuance, while Class B investors received 196 million shares, or 9% of the total. Among private funds, the top 10 by number of allocated accounts were all major quantitative firms. Shanghai Yanfu ranked first with 282 allocated accounts, followed by Century Frontier, Jiukun Investment, Shanghai Chengqi and Huanfang Quant with 209, 194, 167 and 153 accounts. Lingjun Investment, Shanghai Jinde and Minghong Investment also each had more than 100 allocated products. According to public information cited in the report, DeepSeek founder Liang Wenfeng controls both Ningbo Huanfang Quantitative and Zhejiang Jiuzhang Asset. The two firms together had 194 private fund products allocated 20.2497 million shares, with a placement value of about 175 million yuan, the largest share among private funds.