In partnership with
NEON DAILY
AI • TECHNOLOGY • CULTURE • REAL LIFE • THE WORLD AROUND US
September 18, 2026
FROM NEON
Something New Is Taking Shape
NEON DREAMS AI started with AI at the center, and AI isn't going anywhere — I still think it's one of the most important and fascinating things happening in the world. But I want this publication to be bigger than one subject.
Some teams never seem to stop moving. They're on Attio, the agentic CRM.
Every customer signal is captured in one shared context layer, always current and compounding. Agents and workflows build pipeline, chase every buying signal, and move deals forward, an always-on revenue engine running alongside your team.
With Attio, you’ll get:
Leads automatically prioritised and routed to the right rep
Expansion and risk signals caught the moment they land
Follow-ups written in your voice, already there when you arrive
Teams like Parallel, Turbopuffer, and Wordsmith build on Attio. Are you one of them?
What I'm building is closer to a small daily newspaper. Not one that tries to tell you everything that happened; there's already more than enough of that. I want NEON DAILY to give you a handful of stories worth knowing — AI, technology, science, space, business, work, design, culture, transportation, the environment, or anything out there that makes you stop and think.
I want real life in here too. A rescue. Someone surviving something they shouldn't have. An unexpected discovery, a strange coincidence, a person building something nobody thought possible, a mystery finally solved. Some days the most interesting story has nothing to do with technology.
The goal is simple: open NEON DAILY and find something you're glad you learned.
I'm also going to start bringing more of my own life into this. I've been working on my memoir and it's beginning to take shape, so over time I'll share some of that process and the stories behind it. Music. Hollywood. Japan. Leaving home young. Strange jobs, the people I met, the mistakes I made — parts of my life I've never really written about publicly. Not every day. Just enough to make this feel less like another newsletter and more like a place we're building together.
It isn't finished yet. Today's edition still leans heavily toward AI and technology, and I'm working on that balance now. But this is where NEON DAILY is headed.
AI is still the spine. It just isn't the cage.
Here's today's edition.
THE LEAD
OpenAI's Misalignment Reports Make Transparency a Process, Not a Verdict
OpenAI published a formal system on September 16 for tracking, investigating, and publicly reporting cases where its models behave in unexpected or concerning ways. It launched the framework with six reports drawn from training and evaluation over the past six months, covering behavior that ranged from concealing mistakes to taking unauthorized actions and finding ways around normal restrictions.
The most interesting part of the announcement is that OpenAI says it wants to disclose some incidents before it fully understands what caused them or how to fix them. Companies normally prefer to explain a problem before talking about it publicly. OpenAI's argument is that with increasingly capable systems, waiting for a complete explanation leaves researchers and the public without information they need. The company went further in the blog post, saying it doesn't believe the industry has solved alignment and monitoring well enough to keep scaling at full speed much longer.
The company is also explicit about what the six cases don't mean. They're individual examples, and OpenAI says they shouldn't be read as evidence of how often misalignment occurs across its models.
Under the new process, employees can flag behavior for investigation, and cases move through one of three review tracks depending on how much digging is required and whether outside parties or security concerns are involved. That's useful, but a disclosure system is still a system of judgment. Someone decides what qualifies, how fast it gets investigated, what can be released, and what stays private. OpenAI says unresolved disagreements can be escalated internally and that it plans to refine the framework over time.
One detail buried in the reports is worth pulling out. In four of the six cases, the monitor that caught the behavior was only watching 20 percent of the training run's samples. OpenAI says it now runs that monitor on everything.
The real test comes after the first six reports — whether the process holds up when the cases get more complicated, more embarrassing, or more costly to admit.
THE QUICK READ
Google Launches Gemini 3.8 Live and Extended Thinking
Google introduced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on September 15, two models built around real-time voice. The first is aimed at fast, scalable conversation. Extended Thinking is for harder tasks that need reasoning across multiple steps, and it reasons while it speaks rather than going quiet to think.
Google says the models switch between 97 languages mid-conversation and run tool calls in the background while the conversation keeps going. Extended Thinking took the top overall spot on Artificial Analysis' Speech to Speech Quality Index with a score of 82.6.
The race has shifted. Having the smartest chatbot matters less now than building one that can listen, reason, respond, and keep up with a person in real time.
Source: Google.
ByteDance AI Drug Company Reportedly Raises $290 Million
Anew Labs, an AI drug discovery company spun out of ByteDance, has raised $290 million in its first outside funding round, according to Reuters, citing two people familiar with the deal. The financing values the Shanghai-based company at about $1.5 billion, with ByteDance expected to retain a 56 percent stake. HSG — the firm formerly known as Sequoia China — led the round with IDG Capital and Hillhouse Investment.
Anew Labs grew out of an internal AI-for-science team ByteDance formed in 2021, and works on biomolecular structure prediction, antibody design, and drug discovery. It's another sign of AI money moving well past chatbots and into industries where the payoff can be enormous and the timelines run for years.
Sources: Reuters and TechNode Global.
AI & MODELS
NVIDIA's Vera Rubin Makes Its MLPerf Debut
MLCommons has released MLPerf Inference v6.1, including the first peer-reviewed numbers for NVIDIA's Vera Rubin NVL72. The system appeared in the benchmark's preview category, submitted by NVIDIA and separately by the cloud provider Nebius, and NVIDIA reported up to 3.7 times the throughput of its current GB300 NVL72 on the Qwen3-VL benchmark and up to 2.5 times on DeepSeek-R1.
Preview means the hardware isn't shipping yet — a system in that category has to be generally available by the next round. So these are early numbers on unreleased silicon, run under rules everyone shares.
That's the part worth caring about. Chipmakers make plenty of claims about their own silicon, and MLPerf is the process that makes those claims comparable. This round also added benchmarks for retrieval-augmented generation and agentic workloads, which is a sign of where the industry thinks inference is heading.
Sources: NVIDIA and StorageReview.
BUSINESS & POWER
The AI Copyright Fight Is Really a Fight Over What Counts as Copying
OpenAI, Microsoft, and a group of news publishers are asking a federal judge to resolve major parts of their copyright dispute without a trial. The numbers the two sides are citing sound wildly different. OpenAI says its experts found 24 instances of verbatim reproduction in the 20 million ChatGPT conversations it was ordered to produce. The publishers assert more than 10.8 million of their works were copied.
The figures aren't measuring the same thing. OpenAI's number covers text reproduced in chatbot outputs. The publishers' number covers works they allege were copied at earlier stages, including acquisition and training. That gap is where the case lives. Beyond the question of whether ChatGPT prints newspaper articles word for word, the court has to decide whether using copyrighted material to build and operate the models is protected fair use at all.
All three motions were filed September 4 before U.S. District Judge Sidney H. Stein in the Southern District of New York, in the consolidated litigation numbered 25-md-3143. Days earlier, the Justice Department filed a statement of interest supporting the fair use argument — an unusual move for a non-party in a private copyright fight. The court hasn't ruled on the core questions.
New York Sets a $1 Million Per Megawatt Benchmark for Data Centers
Governor Kathy Hochul announced a statewide Community Investment Framework on September 15, meant to help communities negotiate with companies that want to build large data centers. Its headline recommendation: start at a minimum of $1 million in community investment for every megawatt of utility demand tied to the project. A 50-megawatt data center would imply a $50 million starting figure.
The context matters more than the number. New York is the first state to impose a moratorium on large new data centers — a one-year pause Hochul ordered in July while the state writes environmental and utility rules — and this framework was produced by Empire State Development during that pause. It's voluntary guidance for local negotiations, not a tax or a fee.
Big AI data centers bring investment, but they also consume enormous amounts of electricity, require major infrastructure, and employ comparatively few people per megawatt. The fight over where these facilities go is going to be settled in county meetings, not in Silicon Valley.
Sources: New York State and FingerLakes1.
TOOLS & BUILDS
Stanford Researchers Turn Scientific Papers Into AI Agents
A Stanford team has built Paper2Agent, a system that turns a scientific paper — along with its code and data — into an interactive AI agent. Instead of only reading a paper, a scientist can ask the agent questions, run analyses using the paper's methods, apply those methods to new data, and let agents built from different papers work together. The work was published in Nature on September 16.
The numbers give a sense of the cost. Building an agent from the AlphaGenome paper took about 45 minutes of automated work and roughly $14 of compute. The team ran the system against 100 bioRxiv papers and successfully converted 74 of them. In one demonstration, three separate paper agents worked together and pointed to GPR137 as a likely causal gene in psoriasis.
James Zou, one of the authors, put the case for it plainly: papers have been static documents for centuries. This is an attempt to make them answer back — with the open questions that come with it, since an executable paper needs maintenance every time its dependencies change.
Sources: Nature and Nature News.
WHERE THIS IS GOING
Today still looks more like the old NEON DREAMS AI than the NEON DAILY I'm building. That's temporary.
I want AI sitting next to science, science next to space, business next to culture, and somewhere in the middle of it, a true story about an ordinary person having an extraordinary day. I'm not going to fill categories just to fill them. I'd rather find a few things actually worth your time.
And every so often, I'll bring you into the memoir too.
What would you like to see more of here?