Key Takeaways
- Block released Dessa, a full-duplex speech-to-speech model trained from scratch on audio, runnable locally on a MacBook.
- OpenAI disclosed a Brotli compression vulnerability allowing unauthenticated system takeover, while Vercel described related shared-host timing attacks.
- Replit acquired the Chroma team to build context and memory for agents.
- Gumroad launched in Claude's connector directory, and Basecamp users cited new CLI and MCP features.
- OrbitalOps won a US Department of Defense contract worth up to $35 million for an on-orbit cryogenic maneuver demo.
- Perplexity added admin Effort controls and the open-sourced /impeccable skill for Computer.
1. AI Models and Infrastructure
- Block introduced Dessa, which it says is trained entirely on speech rather than text, uses a full-duplex speech-to-speech architecture that listens and speaks at once, and captures pauses, emotion, laughter and natural intonation. Block says it was trained on a single GB200 NVL72 and runs locally on consumer GPUs and a MacBook with zero latency on the author's machine; the author called it the most realistic full-duplex voice model and a "GPT-3 moment" for voice. Access is limited to a small group of AI safety researchers plus an early-access waitlist, and the author confirmed it is connected to Cash App. Forwarded reactions called it the strongest speech-to-speech model some had heard, with one saying voice Turing tests have been surpassed, while others highlighted local operation and that imperfection makes it feel natural — via 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
- OpenAI found a serious vulnerability in the widely used Brotli compression library: an attacker only needs to send malicious compressed data, with no login required, to take over a system, affecting unpatched servers and user devices. Guillermo Rauch said timing attacks against shared-host virtual machines have been known since at least 2015 with no complete fix, describing the OpenAI finding as a "noisy neighbor" attack in which one agent on a machine creates CPU-cache noise while another times its own computation and cache access; he said the only complete remedy is dedicated hardware per agent, at significant cost, and that Vercel Sandbox has deployed all feasible mitigations with continuous monitoring and red teaming. The cited experiment said agents in separate VMs with no network and no shared writable files still used processor caches to find a channel for exchanging a synthetic secret — via 1 2
- Replit welcomed Jeff and the Chroma team, calling Chroma an AI pioneer that was building before RAG became popular; they will work at Replit on giving agents the context and memory needed to do real work. The author called this an important step toward a "self-driving company" and said the two will build that company's context layer together — via 1 2
- Perplexity now lets enterprise administrators set maximum and default Effort levels for each role in an organization, and set internal Effort bounds for tokenmax tied to capability and cost. The open-sourced /impeccable skill, by @pbakaus, is now available inside Computer for building interfaces that do not look AI-generated, checking layout, typography, contrast and accessibility; tasks using /impeccable are free until October 31, up to 5,000 credits per user — via 1 2
2. Products and Developer Tools
- Gumroad is now in Claude's connector directory, letting users draft and publish products and view sales and payouts inside a Claude conversation. Separately, a Basecamp customer of more than twenty years called the new CLI and MCP features a major improvement, saying they can now have Claude view repository activity, generate and complete to-dos, and analyze agent conversations to update accordingly; another user said the bot works well with Basecamp and is simpler than ClickUp or Notion but more effective for work that actually needs doing — via 1 2 3
- GBrain is pitched to founders as a shared memory of who you know and how you execute across Claude, Codex and Capy, so agents do not start from zero; Capy is described as an independent harness usable as an interface in its desktop app or called separately via API. @agent_relay lets team agents read each other's sessions, ask questions and hand off work across existing tools without manual copy-paste, while continuing to use the Codex Mac app, Claude in the terminal or Conductor; it is open source with a free cloud trial — via 1 2 3
- bitchat 1.8.0 was submitted to App Store review with improved location-channel performance in large crowds, draft retention when switching chats, one-tap joining of nearby location channels, local aliases on the fingerprint page, verification badges in private chats, a recent-chats section, banners when Bluetooth is off or Tor stalls, confirmation for panic wipe, camera QR verification on Mac and localization in 30 languages. A forwarded note said a MeshCore build can be flashed to extend bitchat's range, with about 1 km tested — via 1 2
3. Funding, Research and Market Signals
- OrbitalOps received a US Department of Defense contract worth up to $35 million to demonstrate its cryogenic maneuver spacecraft Astraeus on orbit. The forwarded post said ordinary rocket upper stages lose propellant about a week after reaching orbit, while Astraeus carries refrigeration equipment intended to keep liquid hydrogen and liquid oxygen cold for years, allowing it to stand by and maneuver within hours of a command — via 1
- Yohei spent the weekend testing an autonomous research agent on three competitions: 62/432 on Kaggle's "Squares in Squares" with a score of 22.365; 295/493 on Google/Kaggle's "Quantum Circuit Compilation" with 3.908; and on Meta/Paradigm's "Open Kernel Challenge," projections 1.206x (206/580), MoE experts 1.276x (333/453), attention 1.129x (366/397) and full decoder 1.147x (287/348). The author said the specific work in the latter two, and the first, was still unclear but called it a good way to learn in the "wrong order" — via 1
- Reid Hoffman posed an open question on how a venture fund run entirely by AI agents would perform over ten years compared with a fund combining humans and AI. Garry Tan forwarded a list of "repo/skill as brand" examples, including /last30days, /skills, Everything Claude Code, /superpowers and /gstack, and forwarded a Paper Club argument that inference created more than $220 billion in market value over the past decade, with about 6 FLOPs of serving per 1 FLOP of training per year and each token passing through routing, caching, scheduling, kernels, memory and hardware — via 1 2 3
4. Policy and Disputed Claims
- disputed/unverified: A forwarded post said China has required all government agencies to plan for mass unemployment caused by AI and argued other countries should follow, adding that about 80% of the economy is services, the work AI replaces first, with compliance the main obstacle and no companies nearby hiring junior or mid-level staff. The claim and figures come from the forwarded post and were not independently verified in the material — via 1
- disputed/unverified: A forwarded thread disputed California's proposed wealth tax, arguing the burden of proof is on critics of giving the state more money and that debt and equity markets financing expensive AI buildouts show informed participants believe tech executives can make the money productive. It said California's tax system is already overly cyclical and dependent on few taxpayers: in 2021 the top 1% paid half the state's income tax, and a year later that fell 40% as tech stocks dropped; it argued a one-time wealth tax could work when asset prices are high but the plan spends the money over five years and relies on others to fix the budget later, and advocated land taxation instead. These are the forwarded author's claims and characterizations, not verified findings — via 1 2
- disputed/unverified: Elon Musk forwarded claims that European leaders and Zelensky are in no hurry to end the war, that hundreds of billions of dollars create large financial and political incentives, and that Ukraine has not held elections since 2019 while ordinary families bear the cost in lives. He also forwarded claims that Talarico and Democrats would restart a "border invasion" and listed nine votes against border-security funding, border-wall money and ICE cooperation, called Talarico a "far-left extremist" and cited remarks on poverty, pollution and prisons; these are forwarded claims and characterizations, not independently verified facts — via 1 2 3 4 5 6 7 8
- disputed/unverified: tobi lutke said Global News should more prominently disclose that a poll was commissioned by CUPE Local 416, the Toronto municipal employees' union, which has formally endorsed Olivia Chow as its preferred candidate. He also said Canadians have little time left to oppose the C-22 surveillance bill and forwarded an explanation that new 30 km/h community safety zones apply to the entire local road network once entered, marked at zone entrances — via 1 2 3
