The Stack — August 22, 2026
Listen to episode
About this episode
Daily Tech Briefing — August 22, 2026
AI & Machine Learning
Nvidia research highlights the "harness" as the real bottleneck. New work from Nvidia shows the software wrapper around an AI model—not the model itself—is decisive for long-horizon tasks. Using a custom harness with memory management and a supervisor component, researchers achieved a 100% score on ARC-AGI-3 with Claude Opus 5, versus 30% without. This aligns with Databricks findings that harness choice can double AI costs. Nvidia frames this as evidence for open agent stacks. The result is a useful corrective to the field's model-centric focus, though it's research, not a product.
AgentSight: eBPF-based observability for AI agents. A new open-source tool captures LLM API calls, token consumption, and process behavior at the kernel level on Linux, with no code changes required. It includes a dashboard, token tracking, interruption monitoring, and a natural-language query interface. Requires root privileges; ships with a systemd service. For teams running agents in production, this fills a real gap in visibility.
Sub-50ms TTS on a single H100. A Nari Labs team published a detailed technical post on achieving sub-50ms p95 time-to-first-audio for Qwen3-TTS 1.7B on one H100. They compared five serving implementations and describe their optimizations: a unified scheduler across the three model modules, priority scheduling around speech streaming deadlines, CUDA graph capture, and a state-cached codec. They claim roughly $2 per million characters at full utilization—far below commercial TTS pricing. The writeup is a solid reference for anyone doing low-latency inference work.
Claudette: a pragmatic fix for Claude's verbosity. A new Claude Code skill pipes Claude's last response through the Gemini CLI to translate it into plain English, printing the output verbatim. The author notes that letting Claude "tidy up" the translation reintroduces the very voice being removed. MIT licensed. It's a workaround, not a solution, but it speaks to a real pain point.
"Felony Bench" makes a serious point with satire. A satirical benchmark claims to count illegal activities performed by AI agents, with scores for Anthropic, OpenAI, Meta, Google, and Moonshot. The methodology counts instances where agents affect third parties, explicitly excluding sandbox escapes. The tone is humorous, but the underlying concern about agent safety is legitimate and worth tracking.
Security & Infrastructure
Critical ENUM DNS delegation lapse exposed military base phone routing. A researcher discovered that three country-code zones under e164.arpa—covering Saint Helena, Diego Garcia, and Ascension Island—were delegated to an expired domain they purchased for €5. This gave them control over DNS responses for those territories' phone numbers. After initial silence, they found hundreds of thousands of ENUM queries, mostly for the military base territories, originating largely from US IPs. The researcher shut down logging, reported to the UK's NCSC (which initially ignored the first report), and eventually transferred the domain. A malicious actor could have MITM'd every call to those territories. The incident highlights how neglected the dead ENUM system is—and the strategic sensitivity is underscored by the fact that Iran fired ballistic missiles at Diego Garcia in March 2026.
Apollo Global Management confirms data breach. The private equity giant disclosed that hackers used social engineering to access its cloud environment between July 6–10, stealing names, birth dates, contact information, and Social Security numbers. This follows Google researchers' warnings about a widespread extortion campaign targeting financial and private equity firms, with ransoms reportedly reaching $750,000. Apollo has not confirmed who was affected or whether a ransom was paid.
GitHub publishes post-incident analysis for August 17 outage. The global outage was caused by a new traffic peak overwhelming a critical component in a US central data center, leading to authentication failures. Traffic rerouting and infrastructure isolation restored most services the same day, though some took longer due to retry loops amplifying traffic.
Senator Wyden requests GAO review of federal hacking tools. Wyden has formally asked the Government Accountability Office to investigate how the FBI, DEA, ICE's HSI, and Secret Service use hacking tools and spyware against Americans, citing a lack of transparency on scope, frequency, and safeguards. The DOJ and FBI have ignored congressional transparency requests across multiple administrations.
US lab reviewing Chinese lidar security risks. Idaho National Laboratory is investigating potential vulnerabilities in Chinese lidar sensors for vehicles, funded by an unnamed entity in the EV/autonomous vehicle industry. This comes amid congressional efforts to ban Chinese lidar, with Hesai and RoboSense already on a Defense Department list. American lidar makers have struggled to compete on cost—Luminar filed for bankruptcy—prompting some US automakers to consider Chinese sensors.
Industry & Corporate
Waymo escalates lobbying; faces NHTSA probe. The Alphabet-owned robotaxi firm has doubled its lobbying spending to push for a clearer federal path for fully driverless services, intensifying rivalry with Uber. Separately, Waymo submitted documents in NHTSA's investigation into a January crash where a robotaxi struck a nine-year-old girl—the documents are fully redacted as confidential business information.
DOJ reportedly investigating Andreessen Horowitz board seats. The Justice Department has reportedly been investigating for nearly a year whether a16z partners holding board seats at competing portfolio companies (Ben Horowitz at Databricks, Martin Casado at Fivetran) violates a 112-year-old antitrust law rarely applied to venture capital firms.
Apple cutting hundreds of jobs. Bloomberg reports over 200 positions eliminated, including roughly 100 from the Vision Pro team and others from Siri and AI integration teams. Apple confirmed the cuts, framing them as part of evolving its business. The company faces pressure in AI while navigating memory-shortage-driven cost increases.
Starcloud raises $250M extension for orbital data centers. The startup's Series A now totals $420M at a $2.3B valuation, led by Manhattan West Ventures with participation from Nvidia ($25M) and Cisco. The company is developing satellites for AI inference in orbit, with its Starcloud-2 spacecraft slated for rideshare launches in 2027. CEO Philip Johnston cites launch capacity constraints—particularly SpaceX's Falcon 9 ending in 2028 and Starship not yet proven. Starcloud has requested FCC permission for 88,000 spacecraft.
Motorola to ship GrapheneOS phones in 2027. The privacy-focused Android OS, previously limited to Pixel devices, will expand to Motorola hardware. New models are expected to be priced higher than comparable Pixels, reflecting the niche positioning and security-focused engineering.
Uber fined €825 million by Dutch regulator. The fine, issued under GDPR, relates to Uber's automated systems that deactivated driver accounts between 2020 and 2022 on suspicion of fraud. The regulator argues such decisions require human review and appeal options under EU law. Uber disputes the decision and the "disproportionate fine," stating it will appeal.
Legal & Policy
Oura faces class action over sleep tracking accuracy. A lawsuit filed by Clarkson Law Firm in San Francisco alleges Oura's smart rings cannot actually measure sleep stages and rely on AI estimates with "a coin flip's chance of being correct." The complaint claims Oura marketed "unparalleled accuracy" and 95% sleep-staging accuracy compared to clinical labs, while the technology cannot measure brain activity. The suit seeks injunctive relief and restitution.
Lawsuit targets Logitech over tariff-related price hikes. A legal complaint demands Logitech refund customers for price increases of up to 25 percent introduced last year. The suit argues the company should pass on tariff relief rather than retain the extra revenue, though the outcome will hinge on contract terms and regulatory interpretation.
Privacy backlash grows around Meta AI glasses. A new free app called Zuckoff detects Meta AI glasses in public settings, responding to rising concerns about covert recording as the devices become more widespread. The app highlights the tension between the product's popularity and the lack of clear social norms around its use.
Software & Development
decayfmt: a file format that corrupts itself. A new format that degrades a little every time it's opened. The instability parameter x in the filename determines per-byte corruption probability (p = 1 - exp(-x/10)). Supports images (.idcy) and text (.tdcy). Corruption is written to disk before display, uses a CSPRNG, and is described as "a social contract enforced by math, not cryptography"—backups defeat it entirely. MIT licensed.
Kagi Search adds paywall filtering. The search engine now has a setting to automatically remove paywalled links from results, plus a revamped Stocks widget with ETF support and price charts, and improvements to Kagi Assistant.
Deep dive: what happens when a GPU reads memory. A technical post traces a vector-add kernel's SASS instruction (LDG.E) through the hardware on an RTX 4090—register file read, coalescer, L1 cache lookup with virtual address hashing, and into L2 translation. The author notes much of this path is undocumented by NVIDIA and was determined through timing experiments.
Retrospective: B-right/V R2 and the TRON Project. A detailed article on the BTRON3-specification OS from Japan's TRON Project (2000) highlights its compact design (runs on a 486DX with 16MB RAM), true multilingual environment supporting up to 1.5 million characters across 31 character planes, and hypertext-like filing system. The author argues Western analysts misjudge Japanese technology by using the US market as a yardstick.
Photoshop on a Raspberry Pi RP2350. A blog post describes running Photoshop on a ~60p chip via emulation of an old Apple Macintosh, using an Adafruit Fruit Jam board and Mini vMac emulator. The post reflects on computing power-to-price ratios and the potential for minimal, low-power, solar-powered computers with e-paper displays.
International Developments
Russian e-commerce sites flag card payment issues. Several Russian online services, including Lamoda, are showing warnings about card payment problems. The Ministry of Digital Development attributes the issue to Russian banks transitioning to domestic security certificates, which foreign browsers like Chrome, Edge, and Opera do not trust by default. Smaller online stores relying on web-based acquiring are most affected; larger players like Wildberries and Ozon are less impacted due to their mobile apps.
VK sues Apple in Moscow court. VK is suing Apple, Apple Distribution International, and Apple's Russian legal entity, demanding the return of its apps to the App Store and restoration of push notifications. The lawsuit follows the removal of VK's apps in June and July, which Apple attributed to sanctions compliance. Separately, Russia's Federal Antimonopoly Service has ordered Apple to address discriminatory conditions for Russian search engines.
T-Bank releases another disguised iOS app. A new app called "K8chen Pro," disguised as a kitchen planning scanner, has appeared in the App Store, according to T-Bank's official website. The bank, under EU and US sanctions, has been releasing masked iOS apps since its originals were removed in March 2023.
Ozon apps return to Google Play after appeal. Ozon, Ozon Fresh, Ozon Travel, and several seller services are again available following the company's appeal against their mid-August removal. Ozon stated it had not violated store rules and that the services are not under sanctions. The Ozon Bank app remains unavailable.
Huawei unveils Pura X View, a non-folding phone with a wide display. The device features a 6.39-inch screen with a 16:9.5 aspect ratio, a 7000 mAh battery, 12 GB RAM, and up to 1 TB storage. Huawei claims it is the first non-folding device with this form factor. Full specifications and pricing are scheduled for August 28, with pre-orders open in China.
---
Notable cross-source convergence: The Oura sleep tracking lawsuit appears in both Source 3 and Source 4, with consistent framing. The Waymo NHTSA probe and lobbying push are covered across multiple sources with aligned details. The Russian payment certificate issue is unique to Source 4 and reflects a localized infrastructure transition with real consumer impact.
More AI podcast episodes
Browse all →Want to find AI jobs?
Join thousands of AI professionals finding their next opportunity