Josh Langner retweeted
Everyone in cybersecurity has AI now.
So who wins?
A Japanese battle in 1575 has the answer:
The battle of Nagashino was fought between the Takeda clan and the allied armies of Oda Nobunaga and Tokugawa Ieyasu.
The Takeda were the most famous fighting clan in Japan, and their cavalry was thought to be invincible. Oda Nobunaga had less prestige, but he treated warfare as a system to be optimized.
Guns had been in Japan for over 30 years, and both sides carried them. The Takeda treated theirs as an add-on and fought the way they always had, riding straight at the enemy to break their line.
Nobunaga, on the other hand, built his entire plan around the gun. He picked a spot behind a small river the Takeda would have to cross, put up wooden fences, and lined them with more gunmen than the Takeda had ever faced.
The Takeda charged and were destroyed. Many of their top generals died that day, and the clan never recovered.
It didn’t matter that the Takeda had guns too, or that they had the most feared cavalry in Japan.
They didn’t rebuild their strategy around the gun, and it cost them everything.
The situation in cyber security today is akin to the battle of Nagashino.
Everyone has access to AI. Attackers have already rebuilt their operations around it. One hacker breached 9 Mexican government agencies with AI running most of the attack.
The bad guys are getting faster. It’s up to defenders whether they do the same, and use AI to find vulnerabilities in their own code before attackers find them first.
Josh Langner retweeted
We’ve heard again and again that enterprises don’t want to pour all their valuable IP, data and tacit knowledge into someone else’s model. They want control of their own weights.
Every single one needs the ability to function independently. This incident highlights a serious issue we have with highly consolidated critical IT infrastructure.
We need to start rethinking our systems.
Alright, three segments in an hour, that was fun. I'm off CNN for tonight.
My money is botched updates on FAA Telecommunications Infrastructure (FTI) which runs over 4400 FAA sites which is the main method for telecommunications based on pilot testimonial.
Seems the most likely.
TLDR: Flights will suck for the next week minimally.
1
44
Josh Langner retweeted
I'll be on @CNN around 7PM ET today talking about the FAA outage.
First - no-one knows what happened, not even the FAA, the FBI, or anyone else. These things take time - so anything is speculation and not conjecture.
Second - the FAA systems are super old dating back to the 1980s with STARS deployed in the 2000s to 2015.
The new AI systems being deployed
The GAO has every year stated the FAA is in a dire straight for reliability (and security). FAA’s 2023 assessment classified 105 of 138 systems, about 76%, as unsustainable or potentially unsustainable.
As of September 1, 2026, FAA said it had replaced over 64% of its copper-based TDM telecommunications infrastructure, leaving roughly a third awaiting replacement.
The modernization plan also illustrates how much equipment needs work. It covers 450 legacy voice switches, digital radio conversion at 1,581 sites, 612 radar replacements, and 53 surface movement radar replacements. Those are program totals, not counts of what remains today. Most of that work targets completion in 2028.
A new system called SMART (AI-based) has been used in limited capacity. SMART began limited operations around Washington, D.C., in September. On November 10, FAA announced its expansion to airspace managed by all 22 air route traffic control centers.
If cyber-related (which again no-one knows) - more modern systems require substantially more interconnected systems that if not properly introduced, can increase attack surface by allowing multiple communications in through third parties, other networks, and less network segmentation versus the older system which was highly segmented from public Internet.
My predictions (purely speculation) - if not cyber related, flights resume tomorrow. Will cause massive flight disruptions for a week or so.
If cyber related - flights probably resume on Monday or later.
💯… but to be clear we’re talking about AI “skills” not people skills 😁
Realization: Avoid personal skills. They lead to process fragmentation and wide variations in developer productivity.
Instead, I strive to share skills across the team.
Shared skills benefits:
- Solve problems using a consistent process
- We can iterate and improve our shared skills over time.
- Fosters team-wide learning
- Avoids "works on my machine"
1
76
Josh Langner retweeted
The future will still have folks who do not want to use AI to build their own tools, and those folks will be soon drowning in a sea of generic AI-made products and platforms. It will be key to find and follow 'producers with taste', people who build with a strong point of view.
Josh Langner retweeted
"Materialize new views as you click "intents", or upsert visual changes in your existing interface. Software starts to feel magical."
Now we can add "fake doors" to product UIs and let the model generate the rest 👨🔬
Introducing an RFC for intent-driven replay-able generative UI.
Generate an interface. Click a button that intends to take you to a new page, generate that page on the fly. Coming back to that page is instant.
intent-ui.com
Couple of interesting topics to dive into:
Infinite Software
Rather than hitting dead ends in your favorite software tool with wishes it could do X, Y or Z... what if you could just tell it to add the functionality and it appears inline with the rest of the content. You get to mutate and mold the experience to your expectations. Mix capabilities from your favorite tools (e.g. Add a "create ticket" button next to a list of Github repos that when clicked automatically create a Linear task).
Intents
Today software is usually made as a whole, meaning every path needs to be well thought out before we can ship it. Most users don't use 100% of the surface area either and everyone seems to use it differently. With intents you can place a button on the screen with an intent, such as "Open a new page with details about this musical artist", and without having to have generated that view ever before we can ship the software. When a user clicks that button that screen will be generated on the fly!
Materialize new views as you click "intents", or upsert visual changes in your existing interface. Software starts to feel magical.
Persistence
An important component of this RFC is the ability to persist changes over reloads, and consistently rebuild the entire UI. Page reloads reference the JSON page structure to lay it out. Tool calls and API requests are made on load to populate latest data into the slots in the UI – but we only make the calls we need. LLM calls stay totally out of the loop on every refresh, only used to amend experiences.
Data Connections
Bring any MCP server or OAuth service as a configuration and access their data. Write operations require human approval, read operations drive the visible UI generation. Ideal world would be you set up all your connections one time and use them anywhere you go. Github, Linear, Cloudflare, any service... you name it.
UI Library Agnostic
The RFC does not require use of any specific UI library so you're encouraged to bring your own. Bind specific UI components to a catalog. For example I bind "Modal" to "LayerDialog" from Kumo (kumo-ui.com)
----
Imagine not being limited by the developers vision and being able to expand upon existing products however you want. Add new functionality on the go. Today the concept works on green field projects, but will continue exploring what a product might look like if this option was given to users to mutate their UI.
The RFC specifies Intents and Persistence (IAP): a JSON document format and the host behavior that lets a model-generated application be stored, reloaded and replayed predictably.
Give it a read and let me know what you think!
Josh Langner retweeted
Absolutely incredible: MIT researchers say GPT-6 Astra improved a small Qwen model’s math score from 7.9% to 48.1% by directly editing just 0.016% of its weights, without gradient-based training.
The result points toward AI systems that can diagnose and improve other models directly, a potential step toward self-improving AI.
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…
Josh Langner retweeted
BREAKING: For the first time in American history, the Department of Energy has approved a commercial nuclear microreactor's safety analysis for full-power operations.
LA-based Radiant secured the approval, with testing planned at Idaho National Laboratory.
Josh Langner retweeted
running NVIDIA Nemotron 3.5 Transcribe (30B-A4B) on a single RTX 3060. 12GB VRAM, 16GB RAM. (config below)
90 seconds of three people talking over each other, done in 27s. speaker labels and word timestamps in one pass. 1 hour context.
His Girl Friday (1940) is famous for the fastest overlapping dialogue ever filmed, so i threw it at it.
stock NeMo-Speech.cpp can’t offload experts to cpu yet, so i added a small patch: n_cpu_moe (like llama.cpp --n-cpu-moe) plus a higher output budget, because the default cuts dense crosstalk off halfway.
config:
nemo-speech serve
--asr-model NVIDIA-Nemotron-3.5-Transcribe-30B-A4B-AD-IQ4_NL.gguf
--asr.salm.n_cpu_moe 40
--asr.salm.max_sequences 1
--asr.salm.n_ctx_per_sequence 73728 --asr.salm.max_context 73728
--asr.salm.n_batch 1024 --asr.salm.encoder_max_batch 4
env: NEMO_SALM_LOAD_MODE=mmap NEMO_SALM_OUT_TPS=12
Meet NVIDIA Nemotron 3.5 Transcribe 🎙️
An open-weight, promptable 30B MoE model with ~4B active parameters that brings transcription in 40 languages, speaker attribution, translation, contextual prompting and audio understanding into one checkpoint.
Watch it keep track of who said what when everyone’s talking at once.
1/ 🧵
Josh Langner retweeted
Picard taught me more about leadership than any MBA could.
He never raised his voice to be heard. He listened before he decided. He let Riker, Data, Worf argue him out of positions, and then owned the call completely. When he was wrong, he said so in front of the crew. When he was right, he didn't gloat.
And when he had to stand alone — against Starfleet, against the Borg, against a Cardassian with four lights — he stood.
That is the whole job. Build people who will tell you the truth, then carry the weight yourself.
Patrick Stewart played it so well that a generation of us mistook a character for a manual. Turns out the manual was fine.
A great actor, and a Yorkshire kid who started as an obituary writer. Today the obituary is his.
Make it so, Sir Patrick. Rest well.
Josh Langner retweeted
For the past year I’ve been heads down, iterating with a small team, building what I believe will be my life’s work.
picardo.health/letter
While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected.
Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions.
Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent.
But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models?
We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem.
Read the full report: cohere.com/blog/concerning-b…
Josh Langner retweeted
The market doesn’t seem to understand that a feature is not a sustainable competitive advantage.
It never was, but now more than ever.
Josh Langner retweeted
We’re like one hack away from me finally being able to get my medical records easily
Josh Langner retweeted
✨ Every big idea starts somewhere.
Ada Lovelace imagined possibilities for computing far beyond calculations.
This week, we celebrate her vision and everyone who continues to shape the future of technology.
💙 Happy coding!
#AdaLovelaceDay #VSCode




























