Episode 2: RIP Bots?
October 29, 2025
Listen
Notes
Welcome to The Silicon Diet - your digest on the latest happenings in AI, fundraises in the Bay Area, and insights into a variety of AI tools.
About Your Hosts
Abhirup - Co-founder and Head of Innovation at Sainapse, an AI customer support company. Been working in the space for about a year and a half, doing his own thing for 2-2.5 years. Based in SF and passionate about AI.
Adi - Regular guy really into AI, cooking up products that will be launching soon. Also based in SF.
RIP Bots? Five shifts remaking AI right now
October 28, 2025
TL;DR
- OpenAI’s hardware play: custom accelerators with Broadcom (10 GW) + a 6 GW AMD/Oracle build-out—real moves to diversify beyond NVIDIA. (OpenAI)
- Chat-first commerce arrives: Apps in ChatGPT, AgentKit, and “Instant Checkout” land alongside Walmart and PayPal integrations. (OpenAI)
- Personal AI compute goes mainstream: NVIDIA’s $3,999 “DGX Spark” puts a Grace‑Blackwell desk-side box within reach for builders. (NVIDIA)
- Autonomy gets spicier: Tesla’s FSD v14 adds aggressive driving profiles and end‑to‑end upgrades; Waymo doubles down on scale and vision-driven research. (Contrary Research)
- Cheaper, faster models: Anthropic’s Claude Haiku 4.5 hits near‑Sonnet‑4 quality at ~⅓ the cost and ~2× the speed—great for coding and retrieval-heavy workflows. (Anthropic)
Deep dive #1 — OpenAI’s great unbundling of compute
OpenAI announced a multi‑year collaboration with Broadcom to co‑develop and deploy 10 gigawatts of OpenAI‑designed AI accelerator racks (Ethernet end‑to‑end), with initial deployments starting in H2 2026 and rolling through 2029. The partnership lets OpenAI embed model insights directly into silicon—classic vertical co‑design. (OpenAI)
In parallel, AMD, OpenAI and Oracle outlined a plan to deploy ~6 GW of Instinct MI450 compute across Oracle Cloud Infrastructure starting in H2 2026—another big bet on multi‑vendor hardware and cloud footprint. (YouTube)
Why it matters:
- Supply & bargaining power: Custom accelerators and second/third suppliers reduce single‑vendor exposure. (OpenAI)
- Networking choices: The explicit Ethernet scale‑up/scale‑out stance signals competition with proprietary interconnect stacks. (OpenAI)
- Energy & siting: Multi‑GW deployments imply multi‑year energy and real‑estate planning—expect siting near abundant power and favorable permitting. (See the timelines and scale in the announcements.) (OpenAI)
What to watch next: tape‑outs in 2026; early rack pilots; AMD MI450 availability; Oracle’s data‑center build cadence. (YouTube)
Deep dive #2 — Chat becomes the app store: Apps, Agents & Checkout
OpenAI rolled out Apps in ChatGPT plus a preview Apps SDK—you can now build chat‑native apps users “open” by just talking to them. In tandem, AgentKit adds the production scaffolding (workflows, evals, guardrails) to ship agents. (OpenAI)
The ecosystem moves fast:
- Instant Checkout (the Agentic Commerce Protocol implementation) is open‑sourced and in ChatGPT—Stripe is a launch partner. (OpenAI)
- Walmart turned on shopping in ChatGPT that can book delivery or pickup (and even tap store inventory logic). (Walmart Corporate News and Information)
- PayPal unveiled ChatGPT support so users can send/receive money inside chats. (The Verge)
Why it matters:
The UX shifts from app‑hopping to agentic flows (discover, decide, pay) in one dialog. This compresses funnels, pushes API surface (Apps SDK + AgentKit) to the foreground, and sets a new bar for trust + citations in transactional experiences. (OpenAI)
Builder’s corner (quick start):
- Prototype with Apps SDK (conversation UX) → 2) Wire actions & tools via AgentKit (guardrails + evals) → 3) Plug in commerce with Instant Checkout patterns. (OpenAI)
Deep dive #3 — The $3,999 desk‑side “supercomputer” (what it is / isn’t)
NVIDIA’s DGX Spark (Founder’s Edition) is live at $3,999: GB10 Grace‑Blackwell superchip, “1 PFLOPS FP4” AI perf, 128 GB unified memory, 4 TB NVMe, and a surprisingly compact form factor. It’s pitched as a personal AI supercomputer for local dev, inference, and small‑scale fine‑tuning. (NVIDIA)
Where it shines
-
Ultra‑low‑latency local inference; prototyping 7B–30B‑class models (especially with quantization and parameter‑efficient finetuning); privacy‑sensitive workflows. (NVIDIA)
Where to temper expectations
-
It won’t replace multi‑rack training for frontier models; memory/bandwidth ceilings and thermals remain real. Treat it as a powerful dev box / edge inference node, not a data‑center. (See NVIDIA’s positioning and specs.) (NVIDIA)
Deep dive #4 — Autonomy: Tesla’s “Mad Max” returns; Waymo scales and explores vision
Tesla FSD v14 made a bigger leap than usual: new driving profiles (Sloth → Standard → Hurry → Mad Max), end‑to‑end network upgrades, and more capable end‑of‑trip behaviors (parking, complex lots, and yes—drive‑thrus in tests). It remains Level 2 (supervised). (Contrary Research)
Waymo, meanwhile, continues to scale (>4.6M rider‑only trips to date) and is publishing research on vision‑based end‑to‑end stacks—even as production vehicles still run multi‑sensor (camera + lidar + radar). (The Road to Autonomy)
Takeaway: End‑to‑end learning + aggressive planning looks more human‑like, but policy + safety cases decide timelines. The sensor debate is narrowing: everyone is leaning harder into learning, with sensing choices driven by operational design domain and risk posture. (Waymo)
Deep dive #5 — The “good‑and‑cheap” wave: Claude Haiku 4.5
Anthropic’s Claude 4.5 family landed with notable upgrades to Haiku 4.5: Anthropic says it delivers comparable quality to Sonnet 4 at roughly one‑third the cost and ~2× the speed, with strong coding performance (e.g., SWE‑bench Verified 73.3%). For agentic coding flows that need lots of fast retrieval and refactoring, that cost‑latency curve is compelling. (Anthropic)
Quick hits we mentioned
- Figure’s new humanoid — Figure 03. Purpose‑built for scale; initial production targets ~12,000/year, with a roadmap aimed at mass manufacturing. Launch video + tech brief are up.
- Dreamforce 2025 recap. Salesforce welcomed ~50,000 attendees, projected $130M local economic impact, and made Agentforce 360 generally available (with partner access to OpenAI, Gemini, and Anthropic). (Salesforce)
- Fal.ai: a generative media platform offering 600+ image/video models, serverless GPUs (scale to 1,000s of H100s), or dedicated clusters—handy for creative apps and hackathons. (docs.fal.ai)
- Sora 2: OpenAI’s second‑gen text‑to‑video is now documented in the API + research pages; keep an eye on evolving safety & content policy around celebrity likenesses and estates. (OpenAI)
Links & references (everything we talked about)
OpenAI hardware & partnerships
- OpenAI × Broadcom — 10 GW custom accelerators (H2’26–2029). (OpenAI)
- AMD × OpenAI × Oracle — ~6 GW MI450 on OCI starting H2’26. (YouTube)
Agentic apps & commerce
- Apps in ChatGPT (Apps SDK). (OpenAI)
- AgentKit for production agents. (OpenAI)
- Agentic Commerce Protocol / Instant Checkout (OpenAI blog + Reuters). (OpenAI)
- Walmart shopping in ChatGPT. (Walmart Corporate News and Information)
- PayPal inside ChatGPT (money in chat). (The Verge)
Models
- GPT‑5 Pro available in the API (DevDay overview). (OpenAI)
- Sora 2 is here (OpenAI News). (OpenAI)
- Claude Haiku 4.5 announcement + benchmark notes. (Anthropic)
Robotics & autonomy
- Figure 03 launch brief.
- Tesla FSD v14: profiles, end‑to‑end updates, drive‑thru notes. (Contrary Research)
- Waymo ridership + research toward vision‑centric learning. (The Road to Autonomy)
Infra & events
- NVIDIA DGX Spark product page (specs & price). (NVIDIA)
- Dreamforce 2025 (Agentforce 360 GA; ~$130M local impact; ~50k attendees). (Salesforce Investor Relations)
- Fal.ai developer docs (600+ models; serverless to dedicated clusters). (docs.fal.ai)
Funding & new labs
- General Intuition raises $133.7M seed to build agentic foundation models for games & the real world. (Fortune)
Want this in your inbox weekly? Subscribe at thesilicondiet.com.