AI Engineer community organizer Swyx
Swyx announced that he is returning as a founding partner with Snyk for the second AI Security Summit. He says every partner from 2025 has returned. Compared with a year ago, he is more concerned about rogue agents, breaches, and AI-driven attacks, and argues that security should be central to the AI conversation. One reward of organizing AI Engineer events has been helping others establish high-quality conferences in their own countries and fields. He included a registration link, but this feed item does not specify the event date or location.
OpenAI Codex and ChatGPT team member Thibault Sottiaux
Thibault Sottiaux says future models will be much better at deleting and simplifying code. He did not name a version or release date in that post. A separate sequence concerns quota resets for Pro 500 users: he first acknowledged reports that some users had not received the expected reset, said the team was investigating, and promised to make up for it. He subsequently reported that everything was fixed and remarked on the unexpectedly large number of Pro 500 users in the discussion. These posts combine an expectation about future code-maintenance abilities with the progression of one quota issue from investigation to resolution.
AI tutorial and interview creator Peter Yang
Peter Yang complained that an interface element stays on screen after he minimizes the Muse desktop app, making him reluctant to open it again. He also described an attempt to replace a software subscription: he had been paying almost $300 a year for a YouTube research tool, but it had become increasingly complicated, so he asked Claude to build only the core features he actually needed. He says Claude completed that subset in five minutes. Another post invited readers to a free short lesson on building an AI system that takes work off their plate, using ChatGPT, Claude, or another AI harness. The post listed 11:30 AM PST on the day it was posted, said more than 500 people had signed up, and promised a recording to those unable to attend.
Meta Senior Director of AI Madhu Guru
Madhu Guru thinks the experience of reading excellent writing produced entirely by a human is underrated. He particularly enjoys the sense of a human presence in a document; this brief observation concerns the reading experience and does not develop a comparison of AI writing quality.
Anthropic Claude Code team member Thariq
Thariq recommends “You should know” as both a way to keep up with what Claude can do and an example of what mods enable. His emphasis is on personalization: users can make Claude Code their own. The post does not include installation steps or a more detailed feature list.
Y Combinator President and CEO Garry Tan
Garry Tan compares Capy to a standup that runs around the clock for agents. He described an unexpected example of cross-session coordination: when his GBrain collaborator Sina started a wave of fixes, it steered around work Garry was already doing across multiple Capy threads. He stressed that this coordination is model agnostic. For GBrain, he wants agents to become as native and expressive as the web eventually became, serving as personal assistants that continually understand their users. His Jiminy Cricket analogy describes a vision of an always-on helper that thinks ahead and keeps improving.
FPV Ventures partner Nikunj Kothari
Nikunj Kothari shared a Claude Code / Codex prompt for home-automation enthusiasts: inspect devices on the home network, list what else could connect to an agent, and research integration methods. His example setup includes a Hermes agent handling personal Gmail, a thermostat, DoorDash, Amazon, and other tasks. The home has an Apple TV available for HomeKit-related work, but no other bridge running Home Assistant. He suggests adapting the prompt to the reader’s own setup. This is a proposed device-discovery and integration-research task, not a completed inventory or report of automation results.
SPC General Partner Aditya Agarwal
Aditya Agarwal says they had their second portfolio-company IPO, which took place this week, naming Puneet and Sanjay at Moneyview. He expressed pride in partnering with the founders and appreciation for the effort they have put into the company. Such moments, he says, bring both pride and recognition of how long and difficult the founder’s journey is. His reason for persevering is the sense of purpose in creating something new. The post provides no further transaction or valuation details.
Sam Altman at OpenAI
Sam Altman responded to speculation about OpenAI’s relationship with Cerebras, describing it as a close partner with which OpenAI is deeply engaged in pushing the frontiers of speed. He did not announce new deal terms or performance figures in the post. Separately, he called dot his favorite OpenAI product so far. As it learns more about his workflow and style, he says it feels noticeably better each day. He particularly appreciates having it handle tasks he dislikes and that otherwise tend to accumulate into a source of dread.
Anthropic’s official Claude account
The official Claude account shared two examples of interactive creations. One uses Opus 5.5 to make an interactive 3D jet engine that can be cut away and pulled apart. The other uses Sonnet 5.5 to create a pop-up city in which each building is a paper fold drawn on a flat canvas. Each post includes a link to the creation. These are examples supplied by the official account; the feed does not state build times or independent test results.
Claude Blog: Claude for Small Business launches new workflows, integrations, and training programs
This Claude for Small Business update was originally dated September 15 and appears in the current feed batch. The product now contains 43 workflows and adds 27 integrations, including Shopify, Salesforce, TikTok, Atlassian, Zoom, Xero, Gusto, Square, Stripe, and Zapier. Workflows extend from back-office tasks to business growth, including lead generation, inbound inquiries, proposals, and everyday reporting. Claude Blog says the product has been installed more than 900,000 times since its May launch.
The priorities reflect feedback from more than 1,000 business owners in 10 cities during the spring tour, about a third of whom wanted help growing their businesses. The fall program plans free workshops in 10 US cities; more than 150 trained organizations will run over 750 community workshops, and 14 integration partners will offer free webinars. Garrett French of Driller Design Co. describes checking the CRM and prioritizing tasks each morning: “We call it the daily briefing.” These are customer accounts collected by the company. A sample week illustrates a continuous workflow from installation and weekly reporting to after-hours leads, proposals, marketing, and month-end accounting.
No Priors — Frontier Chips for Frontier AI Labs, with Walter Goodwin, Founder/CEO of Fractile
Key takeaway: Faster inference matters beyond snappier chat. It can let long-running agents do more reasoning in the same time, but that requires memory bandwidth, capacity, and cost to improve together.
Fractile founder and CEO Walter Goodwin has focused on fast inference for large models since starting the company in summer 2022. He says its roughly 150-person team covers model workloads, front-end chip design, physical design, back-end implementation, and advanced packaging. Keeping these functions together is intended to shorten handoffs and connect architectural decisions with manufacturing constraints.
The first shift is away from relying solely on SRAM. SRAM offers high bandwidth close to compute, but economically accommodating larger models and long-context state becomes difficult. Fractile is therefore working with memory vendors and foundry partners to obtain extremely high bandwidth from higher-capacity, lower-cost DRAM. Goodwin describes a platform planned to ramp in the second half of the following year; this remains the company’s technical direction and plan.
The second argument is that speed can change task capability. Goodwin compares a snappier chatbot to “faster horses”; his larger interest is shortening long-running agent tasks. In his account, some fast-inference approaches still return to GPUs for long-context processing. If their bandwidth advantage cannot cover that state, the whole task does not accelerate at the same rate.
The third constraint is physical lead time and financial payback. He says even an expedited foundry cycle takes three to five months, and a chip needs a three-to-five-year amortization window. Compressing design therefore does not mean selling a fundamentally new chip every few weeks. It means testing more architectural candidates and selecting platforms worth ramping that can remain useful for years. Placement, routing, and the experiments themselves can remain bottlenecks, and final manufacturing-rule checks cannot simply be removed.
Finally, he thinks greater bandwidth could enable sparser MoE architectures, while customers will still need diverse chip suppliers. A model lab tied to one hardware bet could face a dangerous migration delay if a competitor discovers a better algorithm that works on another platform. For Fractile, continually narrowing the gap between identifying a workload opportunity and delivering at volume matters more than winning a single speed demonstration.