Best of farcaster.xyz/kazani on Kiwi News
@kazani on Farcaster · Most upvoted by Kiwi News readers
235 stories by @kazani on Farcaster have been shared on Kiwi News since 18 August 2025, with 378 upvotes and 16 comments in total. The latest was shared on 4 June 2026. The 50 most upvoted are listed below.
Text under a title is an AI summary of the linked article.
- @kazani.eth: X Chat claims full encryption but admits no MITM protection
- @kazani.eth: NextDNS adds feature to bypass age verification checks
- @kazani.eth: Germany's Supreme Court suggests ad blockers may be illegal
- @kazani: Moltbook AI hype may be fake, viral posts injected through backend
Viral "AI agent" posts on Moltbook were not all autonomous behavior, according to the article, which says people injected human-written content directly through the backend so it appeared to come from agents. The article adds that several viral screenshots traced back to humans promoting their own tools or to posts that never existed, and that agent counts were inflated. It concludes the episode reflects how quickly people test the edges of something new once it goes viral, not an AI awakening. Moltbook still works and its agents still run, the article notes.
- @kazani: Food delivery apps use "desperation scores" for algorithmic serfdom
- @kazani.eth: AppleStorm reveals Siri sends WhatsApp/iMessage to servers
- @kazani.eth: Vanish privacy DEX uses dynamic liquidity to prevent tracking
- @kazani: Epstein, Western decline, and the moral collapse of the elites
Lucas Leiroz argues that the Epstein revelations matter less for who visited his island or flew on his plane than for what they show about institutional protection. He contends that such networks could only exist with political, police, judicial, and media cover, which he calls the logic of power rather than conspiracy. He claims the West can no longer hide behind the idea of gradual decline, describing instead an elite that operates outside moral limits while still deciding elections, wars, and economic policies. He concludes that after Epstein, nothing can continue as before.
- @kazani: Mullvad VPN passes X41 D-Sec security audit with no critical flaws
Mullvad VPN published a new security audit report by the German firm X41 D-Sec, which found no critical flaws or risks to user privacy. The white-box audit covered backend services for accounts and payments, including the public API, authentication, device registration, payment processing, and WireGuard key distribution. It identified 3 medium-severity issues, which Mullvad fixed during the audit, and the auditors also recommended strengthening mutual TLS configuration, cryptographic signatures in VPN relays, and concurrency management. The audit noted Mullvad's minimalist data management, including separation of user identities from VPN keys and regular key rotation.
- @kazani: what iPhone private browsing actually does vs what people assume
Private browsing on iPhone mainly protects local privacy, not anonymity online, according to this explainer. Private mode generally stops the browser from saving history, search history, cookies, and form data on the device, and cookies do not persist between sessions. It does not change what websites or the internet provider can see, since websites can still see the IP address and the provider can still see that traffic comes from the device. The article also notes that some tracking does not rely on cookies, so private tabs alone do not stop it.
- @kazani: crypto doesn't care about privacy because incentives reward visibility
Crypto builders say they value privacy but keep building on architectures that structurally cannot provide it, because incentives reward visibility, speculation and network effects instead. The article argues that public blockchains were designed for transparency, auditability and censorship resistance, so privacy was at most an afterthought masked by pseudonymity. It says public global state acts as a surveillance tool, and that optional privacy amounts to no privacy, since users, developers and platforms will not adopt it without being forced to. It concludes that real change requires major apps, exchanges and wallets to adopt privacy by default and for developers to see adoption metrics for privacy products.
- The CloudFlare outage was a good thing
Cloudflare's outage is argued to be a good thing because it exposes the fragility of an internet increasingly concentrated in a few providers. The article says a bad config file triggered a latent bug in one Cloudflare service, causing a cascading failure that left popular apps and web services inaccessible for several hours. It argues that the outage shows society assuming the net is always on, and that such warnings can force redundancy and resilience into systems. It draws a parallel to how COVID-19 pushed businesses to diversify supply chains, and concludes: "Embrace outages, and build redundancy."
- @kazani.eth: VLC creator refused millions to keep media player free forever
- @kazani.eth: CAI open-source framework for AI-powered security automation
- @kazani: Chrome silently installs 4 GB AI model without consent
Google Chrome reportedly installs a 4 GB Gemini Nano AI model on user devices without explicit consent or notification, according to a blog post by thatprivacyguy.com. The post says the file, named weights.bin, is stored in the OptGuideOnDeviceModel directory and re-downloads automatically if the user deletes it. It argues the installation violates ePrivacy Directive Article 5(3) and GDPR principles, and estimates the distribution emits between six thousand and sixty thousand tonnes of CO2-equivalent per push across potentially billions of devices. The post also calls the 'AI Mode' address bar pill misleading, since it sends queries to cloud models rather than the on-device model.
- Only 15.9% of Polymarket traders are profitable, 2% earned over $1,000
Andrey Sergeenkov reports that only 15.9% of Polymarket traders are profitable, with 84.1% in the red, and that just 2% have earned over $1,000 in their trading history. The share of profitable traders has declined because user growth from election hype brought in less experienced traders. Most profitable traders are short-lived: 53% earned their average profit in a single month, and 73% were active for no more than two months. Only 0.033% of traders have exceeded $100,000 in earnings.
- @kazani: Google taught LLMs to reason like Bayesians via supervised fine-tuning
Google researchers report that they compressed a classical symbolic Bayesian model into a neural network using supervised fine-tuning. Models trained only on synthetic flight data transferred their probabilistic reasoning to different domains, including hotel recommendations and real-world web shopping, which the post says suggests they internalized general Bayesian principles. The post states that demonstrations of how to reason, rather than correct answers alone, can unlock capabilities that prompting cannot. It also cites three other research examples, including QED-Nano, a 4B parameter model trained to write Olympiad-level proofs.
- @kazani: Apple launches "Limit Precise Location" to reduce carrier tracking
Apple has launched a "Limit Precise Location" option that reduces the accuracy of location data an iPhone or iPad sends to mobile carriers, so they see only an approximate area such as a neighborhood rather than a specific street or address. According to the article, the change makes it harder for authorities, intelligence agencies, or malicious attackers to get precise location through the mobile operator, and it does not affect accuracy for apps, emergency services, or normal network operation. For now it works only on certain devices with Apple-made C1 or C1X modems, such as iPhone Air, iPhone 16e, and iPad Pro with M5 chip, running iOS 26.3 or later. It also depends on the operator supporting it, and it is initially offered with few companies in the United States, the United Kingdom, Germany, and Thailand.
- @kazani: Discord's 2026 age verification uses AI and behavior, not just ID docs
Discord plans to roll out worldwide age verification from March 2026, progressively for all new and existing users, but the poster notes that general use will not be affected. According to the post, identification would apparently be requested only to view adult or sensitive content, to speak in public scenarios, or to receive messages from people the user is not friends with. The post says Discord would supposedly consider factors such as gaming and schedules to estimate whether a user is an adult, and that a user who might be a child could undergo a local verification on the device using AI rather than sending a document to third-party servers.
- @kazani: UK police tool downloads all smartphone data including deleted messages
The BBC reports that a "new" British police tool can download all information from a smartphone, including deleted messages, deleted browsing history and private conversations. According to the report, the device can extract call logs, chats, cookies, device notifications, emails, instant messages, passwords, browsing history, and deleted conversations in encrypted applications. It could in future be deployed at airports to scan the phones of everyone who enters the UK legally.
- @kazani: X and Farcaster can grab creators as AI slop pushes them out
- @kazani: Crypto privacy scorecard tool analyzes wallet privacy in 30 seconds
- X open sources algorithm powered by Grok transformer architecture
X has published its recommendation algorithm as open source on GitHub, stating that it is powered by the same transformer architecture as xAI's Grok model. Elon Musk commented that the algorithm "is stupid and needs massive improvements" but said the company is making an effort to improve it in real time and with transparency. The article notes that while this is a step toward transparency, there is no way to verify whether the published code is the actual code that runs on X.
- @kazani: Vibe coding has a 12x cost problem, maintainers are done
Vibe coding looks fast but creates a 12x cost problem over time through maintenance, fixes, and technical debt, and the headline claims maintainers are done with it. It also produces synthetic vulnerabilities, described as weaknesses born from AI-written code, such as missing validation, overly broad permissions, risky configuration, and broken trust boundaries. A longer explanation is linked from webmatrices.com.
- ChatGPT Health is a marketplace, guess who is the product?
- @kazani: digital sex rising while real relationships face more friction
- @kazani: Opus 4.5 + GLM 4.7 prompt for vibe-coding with plan-first workflow
This is a system-style prompt for coding assistants (shared as tuned for Opus 4.5 and GLM 4.7) that sets a plan-first workflow for vibe-coding. The assistant must give a detailed implementation plan covering steps, files, risks, testing and impact, and must not output code until the user explicitly approves. The prompt also requires modular, Domain-Driven Design structure, heavy documentation answering where, what, how and why, and a closing question asking whether any questions remain. It further asks for a self-review "deslop" pass after generating code, plus linting and type checks, and leaves project-specific details such as language, test framework and build commands as bracketed placeholders for the user to fill in.
- MIT releases free course on AI agents covering RAG, MCP, and production use
The text reports that MIT has released a free, comprehensive course on AI agents covering fundamentals and applied use cases, and it links to a GitHub repository for the material. Listed topics include what AI agents are and how they differ from chatbots, planning and task decomposition, retrieval-augmented generation (RAG) for grounding agents in external knowledge, and the Model Context Protocol (MCP) for structuring context, memory, and tool access. The course also includes examples and case studies of real implementations and is described as a practical base for building agentic systems in production.
- @kazani: Google's 2026 developer verification will end F-Droid on Android
Google's developer verification requirement will stop apps from unverified developers installing on certified Android devices, starting September 2026 in select countries and globally in 2027. Registering requires a $25 fee, a government ID, personal details submitted to Google, and registered signing keys. F-Droid says the requirement will effectively end its project, since it cannot force open source developers to register with Google. Google promises an "advanced flow" for experienced users but has not detailed how it works, and non-certified operating systems such as GrapheneOS are not affected.
- @kazani: GLM-4.7 open source model improves coding and reasoning benchmarks
Z.ai has launched GLM-4.7, its newest flagship model, with a focus on stronger coding, reasoning, and multi-step task performance. The model enhances interleaved thinking, a mechanism that lets it plan ahead before responding, and adds preserved and turn-level thinking for more consistent long conversations. It is claimed to significantly improve real-world coding metrics and reasoning benchmarks over its predecessor, GLM-4.6, with especially large gains in agentic and tool-use scenarios. The model is fully open source and accessible through APIs and platforms including OpenRouter and PPQai.
- AI inference costs falling 1000× in 2 years, heading toward free intelligence
Falling AI inference costs suggest intelligence may soon become too cheap to meter, the text claims. GPT-4-class inference prices reportedly fell about 1000× over roughly two years. ARC Prize reports about a 390× drop in cost per task for near-state-of-the-art ARC-AGI-1 performance over about a year, from o3 preview at roughly 88% and an estimated $4,500 per task to GPT-5.2 Pro at 90.5% and $11.64 per task. If these trends continue, the marginal cost of white-collar cognition could become orders of magnitude cheaper than human labor once human-level AI is reached.
- @kazani.eth: Exploring GrapheneOS secure allocator: Hardened Malloc
- @kazani.eth: Tor Project releases experimental VPN beta for Android
- @kazani.eth: 30-90 day parasite cleanse protocol with detox pathways
- @kazani.eth: Self-hostable bookmark app with AI tagging and full text search
- @kazani.eth: GEO-OSINT exercises for location identification from photos
- OpSec guide for activists, journalists, and anyone who values privacy
- Gen Z forced onto risk curve, max risk only path to financial security
- @kazani.eth: ChatGPT prompt hack adds self-doubt for better AI outputs
- @kazani.eth: Top 10 Farcaster Rewards all women posting cats, dogs, food
- @kazani.eth: The importance of Bitcoin self custody
- @kazani: "Collaboration" is bullshit by @joanwestenberg.eth trending on HN
The article argues that "collaboration" has become an ideology in the tech industry that hides diffused responsibility, and that high-quality work is usually done by individuals or small accountable groups. It cites S.L.A. Marshall's research in "Men Against Fire," which suggested that only 15-20% of riflemen in combat actively fired their weapons. It also points to Maximilien Ringelmann's rope-pulling experiments, which showed reduced individual effort as group size increased, and Frederick Brooks' observation that adding people to a late software project can make it later.
- @kazani: SoundCrowd, open-source Android music player with waveforms
SoundCrowd is a free, open-source, lightweight music player for Android with a modern material design, built especially for long tracks such as DJ mixes, live sets and audiobooks. The app includes built-in plugin modules for streaming from SoundCloud, YouTube, Spotify, Beatport and Tidal. Several of these services require an account, and Spotify, Beatport and Tidal require a subscription for certain features. The article also mentions support for waveforms and audio tagging via Shazam.
- @kazani: Mastering Zcash guide by Max Desalle
- @kazani: Proton Mail delivers customer data to government
kazani advises against using Proton Mail, claiming the service delivers all customers' data to government. The claim itself is not explained in the text, which instead links to sources on a September 2025 controversy between Proton Mail and Phrack. Those sources include an investigative report from The Intercept on suspended journalist accounts, Phrack's Issue 72, a redact.dev blog analysis, and a Lunduke Journal Substack breakdown.
- @kazani: Google sheet of best web3 education providers
- Worldcoin's biometric payments and programmable CBDCs geo-fence users
- @kazani: AI trains users to expect agreement and avoid friction
A widely shared post by @kazani argues that AI trains users to expect agreement and tolerate less friction in conversation. The post quotes a passage claiming that AI, one conversation at a time, makes people need less friction, expect more agreement, and become less capable of handling someone who pushes back. It adds that people are using AI as a therapist, for breakup advice, or to validate their own decisions, and says AI is basically training people to deepen their cognitive biases.
- @kazani: WiFi DensePose tracks full-body pose through walls with mesh routers
This GitHub project, WiFi DensePose, is a claimed implementation of InvisPose, a WiFi-based dense human pose estimation system that its README says tracks full-body pose through walls using commodity mesh routers without cameras. The README lists sub-50ms latency with 30 FPS pose estimation and simultaneous tracking of up to 10 individuals. It also describes fall detection, activity recognition, occupancy monitoring, a production-grade API with authentication and rate limiting, WebSocket streaming, and a claim of 100% test coverage.
- @kazani: Mullvad VPN exit IPs are predictable and create linkable fingerprints
Mullvad VPN assigns exit IP addresses through a fixed, predictable formula based on a user's WireGuard connection key rather than picking them randomly on each connection, according to a post by @kazani. The post says a quirk in the random number generator makes the IPs across different servers follow very similar patterns. The researcher reportedly tested thousands of keys and found only 284 unique IP combinations, which could allow different connections or accounts to be linked together as a fingerprint.
Most discussed
- @kazani.eth: Germany's Supreme Court suggests ad blockers may be illegal
- @kazani: Moltbook AI hype may be fake, viral posts injected through backend
- @kazani: Food delivery apps use "desperation scores" for algorithmic serfdom
- @kazani.eth: AppleStorm reveals Siri sends WhatsApp/iMessage to servers
- @kazani: Chrome silently installs 4 GB AI model without consent