😁 Hello, super humans! The hardest part of leaving Nvidia was never the silicon, it was the software, and DeepSeek just published a big chunk of the replacement. While regulators and labs argue about rogue agents and model theft, a quieter story is rewriting who gets to train frontier models. Let’s dig in.
📰 Quick Signals
- 🧠 AI: OpenAI says it disrupted a distillation campaign that began July 1 and peaked at 16,000 extraction-pattern requests from over 4,000 users, attributing a core portion to people linked to Moonshot AI.
- 🤖 Robotics: The IFR’s World Robotics 2026 report says professional service robot shipments rose 24% to nearly 250,000 units in 2025, with about 7,000 full-size humanoids sold.
- 💻 Programming: IBM Bob, its agentic software development platform, now supports on-premises and air-gapped deployments for sensitive code.
- ⚡ Electronics: Samsung is reportedly asking about $4 per gigabit for HBM4, more than 2.6 times the $1.5 per gigabit of HBM3E, with negotiations in their final stages.
- 📡 Telecom: Nokia completed the sale of its fixed wireless access business to Inseego for 1.9 million Inseego shares (an 11% stake) plus $10 million in cash for integration costs.
The Big Story: DeepSeek open-sources its own CUDA alternative for Huawei chips
If you build or buy AI infrastructure, this matters: the main moat around Nvidia has been CUDA, and a frontier lab just shipped a credible open-source ladder over it.
What happened: DeepSeek open-sourced TileLang plus six software modules for Huawei’s Ascend chips, mirroring earlier releases it made for Nvidia hardware, according to The Next Web’s report of DeepSeek’s announcement on its WeChat account. Huawei supported the work, and the two companies jointly tuned a 128-card Ascend 950 supernode. DeepSeek says TileLang is now a core tool for its research, and that most operators in its V4 model family have a high-performance Ascend implementation.
The details: The release covers the whole path from kernel to cluster. TileLang is a high-level kernel language meant to sit between model code and the hardware, with a simpler programming model than CUDA. On top of it sit compute libraries (DeepGEMM for matrix math, TileKernels for vector and memory-access operators, FlashMLA for sparse attention), DeepEP Ascend for cross-device communication, and DeepSelect for data selection during training, as summarized by Geopolitechs. DeepSeek claims performance is approaching the limits of the underlying hardware, which is a vendor claim we have not seen independently benchmarked.
flowchart TD
A[Model code: V4 operators] --> B[TileLang: high-level kernel language]
B --> C[DeepGEMM / TileKernels / FlashMLA]
B --> D[DeepEP Ascend: cross-device comms]
C --> E[Huawei Ascend 950]
D --> E
E --> F[128-card supernode]Important
Our take: Hardware substitution was always the easy half; the hard half is kernels, communication libraries, and the thousand tiny operators a real model needs. By releasing all of that in the open, DeepSeek lowers the switching cost for anyone willing to bet on Ascend, and it does so at the layer where Nvidia’s advantage has been stickiest. Treat the performance claims as unverified until third parties reproduce them, but if you write kernels for a living, TileLang is worth reading this week.
🗞️ More News
🧠 AI
- Anthropic says GLM-5.3 produced working exploits on 50 of 410 ExploitBench attempts, and that its open weights let attackers strip safeguards for roughly $4,400 in compute.
- An in-depth interview covers OpenAI’s custom inference accelerator, code-named Jalapeño, including architecture, power, and early benchmarks.
- Goodfire argues interpretability is the main bottleneck to safe alignment and calls for urgent investment in debugging tools for models.
- Ideogram released 4.5, an image editing model built for targeted changes that preserve the surrounding content.
- LlamaIndex launched Extract v2.5, aimed at better accuracy and grounding in document processing.
- Vercel-backed Photon raised $4.5 million to put AI agents inside iMessage and WhatsApp.
🤖 Robotics
- China delivered 19,100 humanoids in the first half of 2026, 97% of global shipments, and German suppliers such as Schaeffler, Bosch, and Schunk see an opening in durable joints and hands.
- Qualcomm will acquire PickNik Robotics and says the MoveIt manipulation framework stays open source.
- Agility’s Digit 5 brings more human-like legs, upgraded batteries, and a stronger safety architecture as its SPAC merger awaits completion.
- Unitree’s shares are down 53% from their IPO debut, wiping out roughly $35 billion from peak valuation.
💻 Programming
- JetBrains Air entered early access in 2026.3 EAP builds, bringing multiple agents such as Codex, Copilot, and Claude into JetBrains IDEs without requiring a JetBrains AI subscription.
- The Cloud Security Alliance’s slopsquatting note found nonexistent package names in 19.7% of 2.23 million code samples from 16 models, with 43% of hallucinated names repeating across reruns.
- E2B Embed lets developers package agent sandboxes inside their customers’ own environments.
⚡ Electronics
- AMD’s Lisa Su is reportedly returning to South Korea in October to deepen HBM4 and NPU ties with Samsung.
- A roughly $30 Raspberry Pi Smart Display Module for the CM5 targets kiosks and digital signage.
- Walnut Pi CM2 is a cheaper CM5 alternative built on an Allwinner T527 octa-core with Wi-Fi 6 and a 2 TOPS NPU.
📡 Telecom
- Samsung signed AI-RAN contracts with KT and SK Telecom under Korea’s Hyper AI Network initiative, with pilots in shipyards and petrochemical plants.
- Ericsson and Nokia are hitting early 6G teething troubles and pushback from operators.
- Verizon says it is moving from scripted automation to agentic network operations where AI reasons across domains while operators keep control.
👨💻 Code Corner
Today’s programming signal is slopsquatting: attackers register package names that AI assistants like to invent. Before you pip install something an assistant suggested, check that it exists and is not suspiciously new.
import json
import sys
import urllib.error
import urllib.request
from datetime import datetime, timezone
def check(pkg: str, min_age_days: int = 90) -> str:
url = f"https://pypi.org/pypi/{pkg}/json"
try:
with urllib.request.urlopen(url, timeout=10) as r:
data = json.load(r)
except urllib.error.HTTPError as e:
return "MISSING (possible hallucination)" if e.code == 404 else f"ERROR {e.code}"
uploads = [
f["upload_time_iso_8601"]
for files in data["releases"].values()
for f in files
]
if not uploads:
return "NO FILES (suspicious)"
first = datetime.fromisoformat(min(uploads).replace("Z", "+00:00"))
age = (datetime.now(timezone.utc) - first).days
return f"OK, first upload {age} days ago" + (" (NEW, review it)" if age < min_age_days else "")
for name in sys.argv[1:]:
print(f"{name}: {check(name)}")Tip
Run it as python check_pkgs.py requests some-package-an-llm-suggested. Existence and age are only a first filter: a malicious package can be old, so still pin versions and review new dependencies.
🧰 Toolbox
- JetBrains Air: runs several coding agents inside JetBrains IDEs, in early access in the 2026.3 EAP builds.
- E2B Embed: packages agent sandboxes so they can run inside your customer’s environment.
- Intrinsic Core: Google’s Intrinsic open-sourced ROS-compatible capabilities for building robotic applications.
- Avaota F2: first SBC with the dual-core RISC-V Allwinner V861, 1 TOPS of AI acceleration and 4K camera support.
- LakeShark firmware: turns the LilyGO T-Display P4 into a handheld SDR scanner with RTL-SDR support.
- NXP FRDM-IMXRT1186: crossover MCU board with Cortex-M7 and M33 cores and four Ethernet ports for industrial automation.
🎬 Demo Watch (rotating)
OpenArm 2.0 is an open-source 7-DOF robot arm with quasi-direct-drive joints, bilateral force feedback, an in-hand camera, and a 4.1 kg payload. Bilateral feedback matters because the operator feels what the arm touches, which is exactly what you need to collect good teleoperation data for training robot policies. Hard parts are backdrivable joints that stay safe around people while still lifting real weight. The hype check: a spec sheet is not a demo, so look for long unedited manipulation runs before trusting any payload or accuracy number.
📚 From the Blog
- Turning Pixels Into Something the AI Can Eat: after clean capture and network transport, the camera series arrives at preparing pixels for the model.
- Building Your First Neuron From Scratch: a neural network is a stack of simple, learnable transformations, and here you build the first one by hand; a good primer for why kernels like TileLang’s matter.
- The Network Behind the Cameras: the unglamorous plumbing that moves video from camera to model without choking the network.
😀 The Bot Says…
CUDA has about 4 million developers; TileLang is currently auditioning for the other 4 million. Please hold your applause until the benchmarks arrive.
That’s all for today! Hit reply and tell us: would you bet a training cluster on a non-CUDA stack?
