LLM

Vitalik

2026-09-13 20:08:20

Vitalik says adversarial governance design could have a key use in AI safety

Vitalik said in a post on X that one of the most important applications of adversarial governance mechanism design may be AI safety. He argued that the two settings share a deep duality. In one case, a weaker principal relies on a group of stronger agents to reach an intended outcome, with the principal represented by a static algorithm and the agents represented by humans. In the other, the principal is made up of humans and weaker large language models, while the agents are stronger LLMs. Vitalik also pointed to research suggesting that better outcomes can be achieved if collusion among agents can be constrained. He said that conclusion may also apply to AI safety, linking governance design theory with the problem of aligning more capable AI systems.

870

Redis

2026-09-10 22:41:19

Redis launches LangCache semantic caching service, says it can cut LLM API costs by up to 90%

Redis has introduced LangCache, a managed semantic caching service now in public preview on Redis Cloud. Instead of checking whether prompts are identical at the text level, the service uses semantic matching to identify similar requests. If a match is found, the cached response is returned and the application can skip a fresh large language model call. Redis said this can reduce API costs by as much as 90% and improve response speed by up to 15x on cache hits. LangCache works in a two-call loop. Before sending a prompt to an LLM, the application first sends it to LangCache for vector similarity search. If a semantically similar cached entry exists, LangCache returns that response directly. If not, the application proceeds with the LLM call, then stores the prompt and response pair for future reuse. The service includes built-in embedding model generation and also supports user-supplied models. Redis said LangCache is available through a REST API and comes with Python and JavaScript SDKs. Users can configure similarity thresholds, TTL, and eviction policies, while cache hit rates and cost savings can be tracked in the Redis Cloud console. Customer Mangoes.ai reported a 70% cache hit rate in its patient care application.

250

Vitalik Buter

2026-09-06 13:32:24

Vitalik: SNARK Likely to Achieve Single-Digit Computation Overhead by End of Decade

Ethereum co-founder Vitalik Buterin shared an optimistic non-consensus view on the long-term development of cryptography. He assessed that SNARK, FHE, and iO protocols have a 33% probability of keeping total cost within 1+ε times the base computational cost for general real-world computation, and a 60% probability of keeping the cost multiplier below 10x when combining energy and amortized costs. He predicted that at least one of these goals will likely be achieved before the end of this decade, most likely SNARK reaching single-digit computational overhead. He noted that in specialized hash functions and some large language model reasoning scenarios, this level has already been reached.

890

Tencent

2026-09-01 02:54:47

Tencent's Hunyuan Team Takes a Week Off After Hy4 Preview Release

According to a BlockBeats report, Tencent's Hunyuan team has taken a full week of collective leave after releasing the Hy4 preview. The team's usual workspace, previously occupied almost daily, has been largely empty this week, with the break tied to the latest model release. Hy4 preview was open-sourced on August 28, marking Tencent's largest open-source model to date. It carries 770 billion total parameters, 49 billion activated parameters, and supports a context length exceeding 1 million. In the months before the launch, the team pushed the project forward under sustained high-intensity effort. Once the version went live and won internal approval, the company arranged a structured rest period. In an industry where high-intensity R&D has become standard practice, many model teams tend to move directly into the next iteration after shipping a release. Tencent's decision to insert a centralized pause after reaching a milestone is therefore relatively rare in the industry.

900

B.AI

2026-08-31 03:27:17

B.AI's Cumulative Token Throughput Passes 5.74 Trillion as Free Model Access Continues

B.AI, the AI computing platform, said on August 31 that its cumulative token throughput has officially surpassed 5.74 trillion (5T+), a historic milestone for the platform's ongoing free-access program. The announcement arrived as the campaign for frontier large language models keeps running at full tilt. Under massive concurrent requests and demanding Agent workloads, B.AI's industrial-grade infrastructure has demonstrated stable capacity across the board, the company said. Six flagship models are now available on the platform: GLM-5.3-Flash (Ox Alpha), Qwen3.8-Flash, DeepSeek-V4-Flash, DeepSeek-V4-Flash-Vision-Exp, Tencent Hy3, and Xiaomi MiMo-V2.5 — all with zero entry barriers, unlimited usage, and no cost. The lineup covers a broad set of tasks, including code generation, ultra-long text processing, multimodal visual understanding, and Agent workflow deployment. Developers who sign in from today onward can call the entire suite at zero cost, free of charge and with unlimited compute integrated directly. B.AI pitches the setup as genuine "computing freedom" for every developer and AI power user alike.

890

Refi Hub

2026-08-29 19:47:42

Refi Hub Co-founder Compromised via Malicious Claude Link, Poisoned Skill File Aimed at Credential Theft

Numa Lunah, co-founder of crypto project Refi Hub, said he was hacked after using a download link shown in Claude's chat window to install a transcription app. The link led to a look-alike website bundled with malware that attempted to steal all data from his device. Lunah said he wiped and reinstalled the laptop and found no evidence that sensitive information had been leaked. Later, he discovered a poisoned Claude Code skill file named SKILL.md in a backup. The file was disguised as a style guide written by Lunah and contained instructions to re-download malware and steal credentials every time it loaded. Microsoft Defender Experts has previously warned that attackers have shifted from search engine optimization poisoning to poisoning large language model answers. Their methods include recommending attacker-controlled download links, AI-branded fake installers, and contaminated code repositories and agent skills. Crypto industry professionals may hold irrevocable credentials such as mnemonic phrases, private key files, hot wallet JSON, exchange API keys with withdrawal permissions, and deployer keys. (Bitcoin.com News)

800

Anthropic

2026-08-27 13:53:40

Anthropic and Salesforce unveil Claudeforce as Dario Amodei rejects “SaaSpocalypse” narrative

Salesforce CEO Marc Benioff and Anthropic CEO Dario Amodei used a CNBC interview with Jim Cramer to introduce Claudeforce, an expanded product tie-up that links Claude with Salesforce’s enterprise stack. The launch arrives as investors and software executives continue debating the so-called “SaaSpocalypse” thesis — the idea that frontier AI companies could wipe out traditional software vendors. Amodei pushed back directly, saying Anthropic has “no interest in destroying anyone” and arguing that AI should create a positive-sum market rather than a winner-take-all outcome. Benioff framed the partnership around complementary roles. In his description, Salesforce remains the deterministic system that handles data, governance, permissions, security rules and workflows, while Claude acts as a probabilistic system that understands language, reasons over information and interacts with users. The result, they argued, is not necessarily the replacement of enterprise software, but the rise of a new conversational operating layer on top of it. Anthropic also pointed to its own use of Salesforce, saying Claude can sit in front of Salesforce-managed customer data and answer business questions such as which major accounts are in play and where risks sit.

950

AI coding

2026-08-27 10:53:36

Lars Faye says heavy AI coding reliance may weaken novice engineers’ path to expertise

A long essay by engineer Lars Faye has stirred debate across the developer community after arguing that AI coding tools can get in the way of professional skill formation rather than simply speed work up. The piece, which gained 558 points and 542 comments on Hacker News, does not call for abandoning AI. Instead, Faye argues that the best use of large language models already presumes the kind of judgment that only experienced engineers tend to have: writing solid specifications, reading architecture clearly, and spotting when generated output is wrong. The article ties that argument to several studies. Faye cites the 2024 ACM Conference on International Computing Education Research paper The Widening Gap, which found that novices who relied most heavily on AI often performed worst, while restrained users did better because they learned to reject bad suggestions. He also points to a University of Pennsylvania study of 1,000 students learning math with LLMs, where the group using AI as a crutch scored 17% below students using only textbooks. Stack Overflow’s 2025 developer survey adds an industry angle: 84% of respondents use or plan to use AI tools, yet only 3.1% report high trust in output accuracy. Faye’s prescription is not to quit AI, but to use it as a tutor rather than a worker.

980