Interesting Finds — 2026-08-25 (5/5)
Meta's Hatch and Watermelon, Perplexity's local-first Portable Computer, Figure's Index dataset, Amazon's automated last mile, Jetson Orin Nano 2, and Keenable's knowledge index.
Continued — final finds flagged 2026-08-25/26.
17. Meta Hatch and Watermelon — consumer agents via WhatsApp
Meta’s Hatch Agent Platform and Watermelon Model (Forkast via Yahoo Finance, Lena Park, Aug 25) reports on Meta’s Hatch — a consumer agent platform planned to run inside WhatsApp and operate across DoorDash, Etsy, Reddit, Yelp, and Microsoft Outlook, with an internal model codenamed Watermelon targeting October 2026.
According to internal documentation cited by The Information, Hatch is inspired by Peter Steinberger’s OpenClaw (viral Nov 2025, banned internally at Meta Feb 2026 for security, Steinberger later left for OpenAI). The report claims Watermelon reaches GPT-5.5 parity on internal benchmarks at roughly 10 times the compute of Muse Spark — unverified, internal only. Meta is reportedly considering a premium tier up to $199.99/month, well above tested Meta One Plus/Premium tiers in Singapore and Guatemala. The broader push follows MSL’s $14.3B acquihire of Alexandr Wang (June 2025) and releases including Muse Spark 1.1, Muse Code, and Muse Glimmer 30B. On the Q2 2026 earnings call (July 29) Zuckerberg framed three buckets — core ads/recommendations, business agents/APIs, and consumer agents for 3.5B users — defending $125-145B capex for 2026. Business monetization is already live: WhatsApp Business Agent at $2.00 per million tokens from Aug 1 (~$0.04-0.05 per exchange).
Finding: Meta testing agentic third-party integration in WhatsApp as precursor to Hatch; Watermelon claimed at 10x prior compute.
Meaning: The technical signal is intra-test conversion of chat into an agent OS. The business signal is willingness to price high-compute consumer agents as a separate SKU rather than bundle them.
Connection: If Hatch holds, the bottleneck shifts from model quality to permission and tool reliability across external sites — the same place OpenClaw hit friction.
Source: Yahoo Finance — Meta’s Hatch Agent Platform and Watermelon Model
18. Perplexity Portable Computer — local-first agent on DGX Spark
Perplexity — A Local-First Agent for Private and Cost-Effective Knowledge Work and the product page Introducing Portable Computer for local-first AI (Aug 25) demo a fully local Computer on NVIDIA DGX Spark — orchestrator, subagent, and harness on-device, with PPLX 27B and Qwen 3 27B running locally, cloud routing only if the user gates it. Summarized via Perplexity’s own research blog and cache: requires Pro/Max subscription plus DGX Spark hardware.
Finding: Orchestrator + subagent + harness entirely on-device; local 27B models for knowledge work with optional cloud.
Meaning: Privacy and cost move from policy to architecture. If accuracy and speed hold against cloud routing, the trade is hardware capex versus per-token spend — and for sensitive work, staying local removes a class of data handling concerns.
Connection: Same vector as Jetson Orin Nano 2 below — smaller, efficient frontier models making edge inference viable — but here applied to knowledge work rather than robotics.
Sources: Perplexity — A Local-First Agent · Introducing Portable Computer
19. Figure Index — a network to build the humanoid dataset
Figure — Index introduces Index: described as “a network of creators building the largest humanoid dataset ever assembled” for Figure 03 / Helix.
Finding: Not a new robot, a dataset flywheel — human demonstrations at scale for a general-purpose humanoid.
Meaning: The constraint for humanoids has been data diversity across tasks and environments, not just model size. A creator network is a way to pay for that diversity directly rather than collect it lab by lab.
Connection: Pairs with the Perplexity and Jetson notes: smaller models become useful when the data they see is broader and closer to deployment. For robotics, that data must be physical.
Source: Figure — Index
20. Amazon — fully automated delivery stations for the last mile
GeekWire — Amazon eyes fully automated delivery stations — Amazon exploring fully automated delivery stations that bring robotics to the last mile. Details were behind bot checks in the fetch, but the direction is unambiguous: after automating fulfillment centers, the next cost center is the station where vans are loaded.
Finding: Station-level automation beyond the warehouse.
Meaning: Last-mile cost is dominated by labor and van time at the handoff. Automating the station reduces dwell and sort errors if it can maintain throughput under variable package mix.
Connection: Boxbot? Backoff Besos indeed — the punchline is that “last mile” keeps moving closer to the door. If stations automate, the remaining human step becomes the doorstep, which is the hardest to automate.
User note: Boxbot? Backoff Besos
Source: GeekWire
21. NVIDIA Jetson Orin Nano 2 — entry edge at 78 TOPS
NVIDIA Newsroom — Jetson Orin Nano 2 (Aug 25, 2026) announces Jetson Orin Nano 2: 78 TOPS AI compute, 8GB memory, 8-core Arm CPU, 2x inference performance of Jetson Orin Nano Super in the same form factor, consuming 40% less power at the same performance in 15W mode. Improved Tensor Cores and higher memory bandwidth. Runs NVIDIA’s open robotics stack plus Jetson agent skills, with optimized local inference for Cosmos, Nemotron, Gemma 4, and Qwen 3. Partners including Cognex, Doosan Bobcat, Matic, and Wing (Alphabet drone delivery) are early evaluators. Ecosystem: AAEON, ADLINK, Advantech, Aetina, Antmicro, Aptiv, Auvidea, AVerMedia, Connect Tech, RidgeRun, Seeed Studio, YUAN and others on carrier boards. Availability H1 2027. Context: more than 3M developers on NVIDIA’s robotics stack.
Finding: Entry edge module that can run current 27B-class language and vision-language models locally at the power and cost point previously reserved for far smaller models.
Meaning: As frontier small/medium models reach prior large-model accuracy, the viable edge moves up the capability ladder without moving up power. That enables real-time reasoning for home robots, vision systems, and inspection/delivery drones where latency to cloud matters.
Connection: Direct complement to Perplexity’s Portable Computer — same thesis (local inference of frontier-small models) applied to physical machines rather than knowledge work.
Source: NVIDIA Newsroom — Jetson Orin Nano 2
22. Keenable — AI-native index of human knowledge — styskin
Andrey Styskin (@styskin) on X (Aug 25): “Today we are announcing @KeenableAI, an AI-native index of the best human knowledge we have, starting with the open web. AI seems to know just about everything until you ask it about something you know deeply. The answer isn’t wrong, but it’s just very average.”
Finding: KeenableAI as an AI-native index, starting with the open web, aimed at surfacing the best human knowledge rather than average retrieval.
Meaning: The stated problem is real — frontier models are broad but shallow where expertise is deep. An index that preserves signal and attribution for high-quality sources would be more useful than another general chatbot if it can rank depth correctly.
Connection: Mirrors Lacuna’s move earlier in this set (733k papers into a traversable graph) — shift work left into precomputed structure so later queries pay graph cost, not parse cost. Different corpus, same principle.
User note: Inevitable but weird
Source: x.com/styskin/status/2092265673041084505