π Hello, super humans! For three years the humanoid-robot conversation has been a spec sheet argument: degrees of freedom, payload, cycle time. Yesterday a county board of supervisors quietly changed the question to something far more boring and far more consequential, namely who signs the permit and where exactly the big red button lives. If you build anything that walks, grips, or carries a lithium pack, this one is about you.
π° Quick Signals
- π§ AI: Zenity closed a $125M Series C led by Norwest to secure and govern autonomous agents that touch production data and tools.
- π€ Robotics: Agility now reports deployment commitments across nine customer facilities and more than 65,000 hours of Digit operation.
- π» Programming: A fresh kernel patch drops the “experimental” framing on Rust for Linux; Rust is now as acceptable as C for new kernel code.
- β‘ Electronics: Europe’s Chips Act 2.0 is still only a proposal, while the Court of Auditors now forecasts an 11.7% EU share of the chip value chain in 2030 against a 20% target.
- π‘ Telecom: The FCC voted unanimously to seek comment on opening roughly 200 MHz of unlicensed spectrum, the 2.4 GHz and 5.8 GHz bands, to satellite direct-to-device links.
The Big Story: San Mateo County wrote the first US rulebook for commercial humanoids
If you have been building humanoids on the assumption that regulation arrives federally, years from now, in a form you can lobby: it did not. It arrived locally, on a Tuesday, from a county board that also handles fire districts and zoning.
What happened: The San Mateo County Board of Supervisors unanimously approved a resolution on Tuesday to regulate commercial humanoid robots, making it one of the first jurisdictions in the United States to do so. The framework, introduced by Supervisor Ray Mueller, requires permits for commercial humanoid deployment, emergency kill switches, fire-safety planning, and tracking of possible job displacement. Permitted robots are also meant to be supervised at all times by a trained, on-site human.
The details: The interesting part is not the AI-safety framing, it is the fire-department framing. Mueller explicitly tied the permitting fees to lithium-ion incidents already seen with e-bikes, laptops and cars, and said the county may need specialized firefighting equipment for robot fires; the permit fee is the funding mechanism. That is a very different regulatory instrument than an algorithmic-accountability law. It reaches your battery chemistry, your thermal-runaway containment, your enclosure, and your service procedures, and it reaches them through a body that can already inspect a building. The permit register doubles as a data-collection tool: the county gets a list of who deployed what, where, and how many jobs moved. Note that this is a resolution setting direction, not a finished ordinance with published fee schedules, so the implementing details are the thing to watch over the next few months.
The reason this bites now rather than in 2030 is that the machines are already on floors. Agility’s Digit has been doing paid robots-as-a-service work at a GXO-run Spanx distribution centre in Georgia, and 1X has taken consumer pre-orders for NEO at $20,000 outright or $499 a month. Meanwhile the standards world has not caught up: humanoid deployment today is stitched together from ISO 10218:2025 and ANSI/A3 R15.06-2025, with ISO 25785-1 still a working draft covering dynamically stable mobile robots. No single standard fully covers a humanoid, so a county filled the gap with a permit.
Here is the chain a “kill switch” requirement actually implies, and why it is a hardware question before it is a software one:
flowchart LR
A["E-stop actuator<br/>(direct-opening contacts)"] --> B["Safety relay<br/>(dual channel, monitored)"]
B --> C["Power contactor<br/>(force-guided contacts)"]
C --> D["Motor / actuator power"]
E["Autonomy stack"] -. "advisory only" .-> B
F["County permit<br/>+ on-site supervisor"] -. "audits" .-> A
B -- "feedback loop" --> B
The point of that diagram is the dotted line. A kill switch that asks the autonomy stack nicely to stop is not a kill switch; it is a feature request. A compliant one removes energy from the actuators through hardware that fails safe when a wire breaks.
Important
Our take: I expected the first real constraint on humanoids to come from a labor board or a federal AI bill. Instead it came from a fire marshal’s budget line, and that is the version of regulation that actually changes engineering, because you cannot argue with a battery. If you are shipping a humanoid, the honest read is that your compliance story now needs a category-rated stop circuit, a documented thermal-runaway response, and a serial-numbered deployment log, and it needs them per county, not per country. Build the safety chain in hardware now, while it is a design decision, rather than in 2027 when it is a retrofit across a fleet.
ποΈ More News
π§ AI
- Hush Security raised a $30M Series A, bringing total funding to $41M, to secure the “non-human workforce” of agents and bots, with Akamai joining as a strategic investor.
- Twitch added an opt-out for letting Amazon train generative models on streamers’ broadcasts, clips, VODs, highlights, chat and images, buried under Security and Privacy in account settings.
- OpenAI paused new signups for its academic-researcher program and moved applicants to a waitlist while it works through more than 13,000 first-wave applications.
- Anthropic switched Claude Code’s auto mode on by default, reducing the number of approval prompts in an AI-assisted coding session.
- Twelve cybersecurity funding deals closed between 15 July and 4 August for roughly $1.09B combined, with agent security and identity governance taking most of it.
- ByteDance shipped Seedance 2.5 on 8 August, the newest entry in a video-generation race that is now measured in weeks rather than quarters.
- Five models landed in the first ten days of August alone, including Muse Spark 1.2 and Qwen Image 3.0 Pro, which makes model selection a per-task cost and privacy decision rather than a standing choice.
π€ Robotics
- 1X opened pre-orders for NEO, pitched as the first consumer-ready home humanoid, at $20,000 outright or $499 a month with a $200 refundable deposit and US deliveries promised for the second half of this year.
- Agility’s Toyota Motor Manufacturing Canada pilot converted into a commercial agreement covering seven Digits in February, alongside more than 100,000 totes moved at the Spanx site.
- August production tallies put Figure 03 past 1,000 units and AgiBot at 15,000 cumulative, which is the context that makes a county permit register plausible rather than theoretical.
- Warehouse robotics rounds have ballooned: the average early-2026 round sits near $42M against roughly $13M in the same window last year, with fleet software and automated storage taking most of the capital.
- A useful compliance primer this week: current guidance for humanoid builders is hazard analysis, force and speed limiting, deterministic fallbacks behind the AI, and secure boot with authenticated updates treated as a safety property.
π» Programming
- Three LTS stable kernels shipped together: 6.18.44, 6.12.103 and 6.6.151, the last of which has now been patched roughly 150 times since October 2023.
- The August TIOBE index left the top ten unchanged, with Python at 18.53% and Rust holding tenth after rising from 1.34% to 1.45% in a month.
- FFmpeg 9.0 and GStreamer 1.28.6 both landed, the latter adding H.266 muxing to its Rust MP4 muxers, which is a quiet but real sign of where media plumbing is heading.
- With Node’s type stripping stabilised in v25.2.0 and TypeScript adoption around 78%, running
.tsfiles without a build step is now the default expectation rather than a party trick.
β‘ Electronics
- The memory squeeze got worse: DRAM spot prices are up roughly 700% over the past year, and HP now puts memory at about 35% of a PC’s bill of materials against 15% to 18% previously.
- Samsung, SK Hynix and Micron, who between them hold over 95% of DRAM output, have pushed capacity toward high-bandwidth memory for accelerators, and SK Hynix’s own CEO has warned the shortage could run past 2030.
- The SIA reported global semiconductor sales up 35.1% from Q1 to Q2 2026, a quarter-on-quarter jump that says more about accelerator demand than about broad recovery.
- ESPHome 2026.4.0 raises the default clock on ESP32, S2, S3 and C5 from 160 MHz to 240 MHz for roughly 34% faster CPU-bound work, adds WIZnet and ENC28J60 Ethernet on RP2040, and moves to LVGL 9, with breaking changes attached.
π‘ Telecom
- AT&T closed its $23B EchoStar spectrum deal, one of the largest single spectrum transfers in recent US history, to feed 5G capacity.
- The FCC also approved an Upper C-band plan freeing 160 MHz for a future wireless auction, keeping the mid-band pipeline alive.
- Operators keep pivoting to compute: Indonesia’s IOH is planning a 1 GW AI factory backed by $800M from Ooredoo, while SK Telecom reports AI data-centre revenue nearly doubling against a 15 GW ambition.
- Airtel is shifting away from fixed wireless access toward fibre, blaming chipset cost inflation, an early warning that FWA margins can evaporate fast in price-sensitive markets.
π¨βπ» Code Corner
A kill switch is only trustworthy if the absence of a signal stops the machine. Here is the pattern in ten lines: a supervisor task that holds a relay energised only while a heartbeat keeps arriving, so a crashed control loop, a hung thread, or a dead process all drop power the same way.
import asyncio, time
HEARTBEAT_TIMEOUT = 0.25 # seconds; tune to your worst-case control period
last_beat = time.monotonic()
def set_enable(state: bool) -> None:
"""Drive the safety-relay coil. Coil de-energised == actuators dead."""
print(f"[relay] {'ENERGISED' if state else 'OPEN (safe state)'}")
async def control_loop():
global last_beat
while True:
# ... do real work: plan, actuate, read sensors ...
last_beat = time.monotonic() # only beat AFTER work succeeded
await asyncio.sleep(0.05)
async def watchdog():
set_enable(True)
while True:
if time.monotonic() - last_beat > HEARTBEAT_TIMEOUT:
set_enable(False) # fail-safe: no beat, no power
raise SystemExit("heartbeat lost, actuators disabled")
await asyncio.sleep(0.01)
async def main():
await asyncio.gather(control_loop(), watchdog())
asyncio.run(main())
Beat after the work, never before, otherwise you are proving the timer is alive rather than the robot.
Tip
This is the software half only. In a rated stop circuit the same idea lives in hardware: the relay coil is held energised through a normally-closed, direct-opening e-stop contact, so a cut wire, a corroded terminal or a lost supply all open the circuit by themselves. Software watchdogs catch logic faults; they cannot catch a broken wire. Build both.
π§° Toolbox
- Humanoid safety standards guide: a readable map of ISO 10218:2025, ANSI/A3 R15.06-2025 and the ISO 25785-1 draft, and what each one does not cover.
- Agility Arc: the cloud fleet-orchestration layer behind the Digit deployments, worth studying if you are designing your own robot fleet manager.
- ESPHome 2026.4.0: free 34% more CPU on your existing ESP32 boards by upgrading, with a short list of breaking changes to read first.
- Octopart: cross-distributor price and stock comparison for safety parts, which matters when the e-stop you specified is on 30-week lead time.
- FFmpeg 9.0: the new major release of the tool that quietly sits under half your pipeline, robotics logging included.
π Component of the Week (rotating)
Omron A22NE-PD / A22NE-P emergency stop pushbutton: the unglamorous part that the San Mateo resolution is really about. It is a 22 mm or 25 mm panel-mount mushroom-head switch with a push-lock, turn-reset action, and its important property is the direct-opening (safe-break) contact: the contact block is mechanically forced open by the plunger, so the circuit breaks even if the contacts have welded shut. Typical blocks are rated around 5 A at 125 VAC with IP65 sealing on the front face, and the family includes a lighted variant plus contact-block combinations you pick per channel. In practice you wire two normally-closed channels into a dual-channel safety relay so a single short cannot mask the stop, which is exactly the chain in the diagram above. Budget on the order of $30 to $60 depending on head size, lighting and contact blocks; the full A22NE datasheet has the contact-arrangement tables.
π From the Blog
- Turning Pixels Into Something the AI Can Eat: episode three of the video-analytics series, on getting camera frames into a shape a model can actually consume.
- Building Your First Neuron From Scratch: the smallest possible learnable unit, built by hand, so the rest of deep learning stops feeling like magic.
- The Network Behind the Cameras: the unglamorous plumbing that moves video without choking the network, and a good companion to today’s story about deploying machines into real buildings.
π The Bot Saysβ¦
Three years of arguing about robot alignment, and the first binding requirement anyone actually wrote down was “please make sure it does not set the warehouse on fire.” Humbling. Correct, but humbling.
That’s all for today! Reply and tell us: does your current robot, drone or rover have a stop that works when the software is dead?


