User says AI art is already passable for average viewers
A user argues AI has won: while details and inconsistencies can be criticized, the overall piece is passable and average viewers don't care.
2026-09-22
All AI Pulse updates published on 2026-09-22.
51 signalsAI makes VC outreach easier, but founders should try harder
AI makes pitching VCs and all outbound easier, but founders should ask whether they would reply to such emails; if not, don't send them.
Good morning greeting post
A multilingual good morning greeting post with no substantive AI news value.
MiMo V2.6 Flash free for a week, now in Go
MiMo V2.6 Flash is free for the next week; both Flash and Pro are also available in Go.
Karpathy on an overlooked Pareto point in LLMs
Karpathy sees such models as a point on the LLM Pareto curve with large latent demand: no thinking, single token, low latency, acceptable intelligence, under-invested due to races for higher intelligence.
Jev: A Zero-Shot Classifier Worth Revisiting
Will Depue reflects on the strong response to Jev, a zero-shot classifier with frontier-level intelligence, and wonders which old ML ideas deserve rescuing.
LangSmith Evals: Agent Quality, Cost, Throughput
Clay evaluates every agent on quality, throughput, and cost. LangChain's Jeff Barg explains how LangSmith evals measure quality and benchmark it against cost and throughput.
Reuters Event: FTC Chair on Tech and Antitrust
Reuters NEXT Newsmaker hosts FTC Chairman Andrew Ferguson in conversation with correspondent Jody Godoy on tech policy and antitrust, Friday, September 25. Registration is open.
Claude Code 2.1.280: Opus 5.5 Now Default
Claude Code 2.1.280 ships 114 CLI changes: default Opus model set to Claude Opus 5.5 (1M context) at $4/$20 per Mtok, plus refined auto-mode behavior for safety review declines.
Inference Fast Enough for Live Emotion Detection
Yohei Nakajima shares that his detection pipeline is now fast enough to power a live emotion detector, suggesting real-time edge AI is nearing practical usability.
Object Detection Latency Down to 0.5 Seconds
Yohei Nakajima reports object detection alone is much faster with roughly 0.5 second latency, making interactions noticeably snappier for lightweight real-time vision apps.
Running Multiple Detectors at Once
Yohei Nakajima demos multiple detectors running simultaneously, further validating the real-time vision system's throughput under multi-task workloads.
a16z Announces Horowitz Andreessen Academy
a16z launches the Horowitz Andreessen Academy, arguing AI now makes it possible to do anything. Courses center on problems too hard to solve without AI.
OpenAI Launches GPT-6 Sol and Luna
OpenAI introduces GPT-6 Sol and Luna, building on GPT-6 Astra to bring its strengths into faster, more affordable models for work at scale, with more efficient caching and inference.
GPT-6 Sol and Luna: Better and 50% Cheaper
GPT-6 Sol and Luna are out, better than 5.6 and 50% cheaper. Luna now costs $0.10 input and $0.50 output per 1M tokens, with output pricing down from $6 in two months.
Binance lists MOONSHOT pre-IPO TradFi perpetual
Binance Futures launched a 24/7 MOONSHOT pre-IPO TradFi perpetual contract for Beijing Moonshot AI Technology, offering leverage up to 20x.
US Central Command revises AI lethal targeting process
US Central Command adjusted AI use and lethal targeting in combat operations, while some experts warn the remaining risk of mistakes is still too high.
Cursor Composer 1 interview tests real project building
Brendan Falk describes a one-hour AI-native coding interview where candidates must build a real medium-sized project using Cursor’s Composer 1 model.
SDSU virtual meetup spotlights agent and physical AI research
Voxel51 is hosting an Oct. 14 virtual meetup on advances in AI at SDSU, covering agentic robotic manipulation and multimodal open-world learning for physical AI.
XPENG IRON demos four human-like interaction capabilities
XPENG showed IRON handling autonomous orientation, dynamic identity memory, seamless multilingual switching, and multi-intelligence Q&A in a real-world interaction test.
Jensen Huang rejects broad US tech ban for China
From CBS Sunday Morning, Jensen Huang again rejected Dario Amodei’s push for a broad ban on selling American tech to China, urging US firms to compete globally instead of exiting.
Haven App Builder turns prompts into team apps
Haven introduced an App Builder where users describe a missing team app and AI writes it, alongside collaborative online/offline work, audit history, and no-code extensibility.
Natural language becomes the 3D workflow control layer
GithubProjects argues natural language is becoming the control layer for 3D work: not just AI-generated assets, but assets agents can understand, reason about, modify, and keep building with.
Meta AI agent fears hit banks insurers and travel shares
Shares of major banks, insurers and online travel agencies fell Tuesday as investors worried tools like Meta’s personal AI agent could disrupt businesses built on consumer inertia.
CactusXR maps store attention with cameras and digital twins
A demo showed shelf attention measurement using only existing store surveillance cameras and a phone-generated digital twin, adding no hardware, via CactusXR.
GPT-6 Sol and Luna — faster, more affordable models with the advances behind Astra’s SOTA per...
GPT-6 Sol and Luna — faster, more affordable models with the advances behind Astra’s SOTA performance in professional work, factuality, coding, computer use, and alignment:
“Hot dog or not?” at 0.4 sec 🌭 jev-style decision model for image [glance-qwen3-vl-4b] now o...
“Hot dog or not?” at 0.4 sec 🌭 jev-style decision model for image [glance-qwen3-vl-4b] now on huggingface and replicate: Demo: https:// glancevlm.replit.app Replicate: https:// replicate.com/untapped/glanc e-qwen3-vl-4b … Hugging Face demo: https:// huggingfa
Here it comes, here it comes, Opus 5.5 is here!
Here it comes, here it comes, Opus 5.5 is here!
https:// upstageplatform.com/faq ?twclid=23d4k1rlphe9b38t8wm5rpq9qi UpStage Platform, (a Sing...
https:// upstageplatform.com/faq ?twclid=23d4k1rlphe9b38t8wm5rpq9qi UpStage Platform, (a Singapore based company) uses AI-powered semantic matching to connect producers and brands directly with creative talent, moving beyond basic keyword searches. Semantic Me
Technology teaches us to challenge limits; some feelings never knew they existed.
Technology teaches us to challenge limits; some feelings never knew they existed. Made on @grok .
INSTEAD OF WATCHING NETFLIX TONIGHT.
INSTEAD OF WATCHING NETFLIX TONIGHT. Spend hours with this. Claude AI FULL COURSE that teaches you how to BUILD and AUTOMATE anything. The people who watch this tonight will wake up tomorrow with a new skill. Watch it and bookmark it now.
ChatGPT changed the game in 2022.
ChatGPT changed the game in 2022. Median revenue four years after founding for startups: 2018: $2 million 2019: $2.3 million 2020: $2.5 million 2021: $2.8 million 2022: $5.6 million More charts:
When Trump and Xi come face-to-face in Washington this week, each president will have his own...
When Trump and Xi come face-to-face in Washington this week, each president will have his own agenda. But one thing is clear: they will have no choice but to cooperate on AI. Tap the link in bio to listen to the full episode of the Big Take Asia podcast https:
We’re working with an independent advisory group of mathematicians to help OpenAI responsibly...
We’re working with an independent advisory group of mathematicians to help OpenAI responsibly share advances in AI and mathematics. The group will advise on how we assess and communicate new mathematical results, uphold academic and professional standards, and
I had Opus 5.5 recreate the city of San Francisco in Unreal Engine.
I had Opus 5.5 recreate the city of San Francisco in Unreal Engine. I powered everything in it (people, pets, cars, traffic, literally everything) with Jev.
Chamath jokes AI jobs remain disputed
Chamath uses a crude joke to say whether AI eliminates jobs is still debated, while anxiety and bad-faith arguments will persist.
Alexandr Wang stresses state of the art
Alexandr Wang posts a blunt state-of-the-art message, without naming a specific model or benchmark.
Google: editable graphs improve long-horizon agents
A Google paper says agents handle long tasks better when workflows live in editable procedure graphs that learn from execution, not buried in chat history.
Salesforce: copying stronger models can backfire
Salesforce AI research finds that after tuning a weaker model’s prompts, tools, and workflow, copying a stronger model can hurt performance; targeted fixes work better.
Building a new IP with Images 2.5
The author says creating a fresh IP with Images 2.5 is now easy, showing BUTTONHOLE BUREAU across characters, spaces, story, and commercial extensions.
OpenClaw 2.0 hackathon starts Sept 22
The @aiworthusing x OpenClaw hackathon with @sodio asks builders to create a multiplayer agent that can do a real startup job, with Mac Minis and an SF trip as prizes.
Clement defends NVIDIA neutrality pledge
ClementDelangue says NVIDIA has clearly committed to keeping the company open, neutral, and silicon agnostic, and promises to prove it through actions.
MiMo-V2.6 details huge RL scaling effort
The author calls MiMo-V2.6 likely one of the largest single RL runs by compute from an open-source model team, staffed by dozens despite scarce compute.
Kimi K3 launches on Amazon Bedrock
Kimi K3 is now available on Amazon Bedrock for coding, document analysis, and extended agent workflows, with explicit prompt caching and AWS controls.
AI software fears lift hardware fundraising
Harj Taggar says hardware fundraising got easier partly because investors feared AI would eat software, and cheaper AI-built hardware could excite them further.
Specialized models offer huge opportunities, says Clement Delangue
Clement Delangue praises Jev: for years, large generalist models dominated AI, but specialized, task- and language-specific models are orders of magnitude cheaper and faster.
Matt Walsh mocks conservatives calling AI “super intelligence”
Matt Walsh says he is already cringing at the thought of conservatives on X dutifully calling AI “super intelligence” simply because Trump told them to.
steipete responds: “It’s definitely a compliment!”
steipete reacts to the claim that Muse and Instinct are both built on his OpenClaw project, saying: “It’s definitely a compliment!”
Muse and Instinct are OpenClaw under the hood
A developer found that Muse and Instinct are both just OpenClaw under the hood, slightly disappointing but a huge compliment to steipete’s work building a kernel for proactive agents.
AI renaming attempts since 1956 share a “glorious fate”
A researcher lists well-meaning attempts since 1956 to rename AI, such as Augmented Intelligence and Ambient Intelligence, quipping that new attempts will meet the same glorious fate.
Yann LeCun adds: “You forgot machine intelligence”
Joining the discussion about historical attempts to rename AI, Yann LeCun replies: “You forgot machine intelligence.”