Mind in the machine ✨Or the other way around, who knows these days?
Denver
Joined May 2023
- Tweets1.3K
- Following340
- Followers16
- Likes1.7K
Morgan Coder retweeted
Current multi-agent recipe: Start a task in a Codex session to initialize a project. Give it all the docs and context it needs. Fork off "Specialist" sessions for areas you want to track progress in separately along with a "Coordinator" session that can manage/report on the specialists.
Astra can do this automatically but manually splitting into focus areas keeps things legible for reporting purposes. I also usually have the Coordinator make a kanban board to keep me updated on tasks.
Morgan Coder retweeted
A critical security issue affecting PHP is being announced in 14 minutes.
Need to immediately patch or take your PHP dependant system offline.
ncsc.gov.uk/news/critical-se…
Morgan Coder retweeted
Neuron interpretability of open source models are now on Neuronpedia
interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had
with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works
already-clean.goodfire.com/
Morgan Coder retweeted
DeepReach (@DeepReach_AI) is building a global network for real-world human data to train physical AI.
Robots and world models need diverse data on how people actually work. DeepReach designs wearable stereo capture devices and places them with local entrepreneurs, who run data collection as their own business in warehouses, workshops, farms, kitchens and repair shops.
DeepReach supplies the hardware, software, quality pipeline, customers and payments. Every new partner adds new environments to the network.
Tim Li and Chris Liu (@drxcliu) launched at YC in Summer 2026. Three months in, partners across 7 countries had collected 500,000+ real-world clips.
@guilleflorvs and I backed them through GR Angel.
xitter.cf/rdominguezibar/status/…
PERSONAL NEWS: After 1 year in stealth @guilleflorvs and I are launching GR Angel
We invest in exceptional founders, put them in front of one of the biggest tech audiences in the world: + 600k followers, + 400k subs & 10M weekly views
Across our personal angel investments and GR Angel, we’ve now backed:
→ A portfolio that has raised $1.2B+
→ Companies including HappyRobot, Replit, Starcloud, Peec ai
→ Alongside investors including a16z, Sequoia, Founders Fund and General Catalyst
→ 1 Exit: Portco went from $10Kmrr to 7M arr in close to 1 year
Our distribution has helped startups get millions of views, increase their ARR, and receive term sheets
Now we’re looking for the most cracked founders worldwide, investing bigger checks, and hiring
you can reach out to us 24 7 👀
Morgan Coder retweeted
‼️ BREAKING: OVHcloud's official CLI was hijacked in a supply chain attack.
A hacker used a developer's GitHub account to plant a binary built to grab secrets in versions v0.11.0 to v0.15.0.
Morgan Coder retweeted
Agents became capable faster than they became trustworthy. They can already book flights, buy things, file paperwork and write code, and the tasks we hand them keep getting bigger.
But before an agent’s actions can be trusted, agentic AI companies need to solve for these things:
1/ Identity: There is no clean way to distinguish an agent’s action from a human’s. Even on your own system, without proper logging, you can’t say with certainty what you did and what the agent did on your behalf. You can’t assign responsibility for an action, or decide whether to allow it, until you know who took it.
2/ Accountability: When an agent acts on your behalf and gets something wrong, it’s unclear who answers for it. It could be you, the company that built the agent, or the company behind the model. Our laws were written for people and for passive tools, and an agent doesn’t fit cleanly into either.
3/ Permission: Even when you’ve authorised an agent, the platform it’s acting on has to allow it too. Most businesses haven’t decided whether they want that. A large part of Amazon’s business is built on people browsing its site, through sponsored listings, reccos and upsells. An agent that goes straight to the purchase skips all of that, which is why Amazon has actively blocked them.
Models will keep improving, but agents will only scale once this plumbing is solved.
Morgan Coder retweeted
> Huawei Ascend 950 NPUs
One of the most interesting things about this release is the support for Huawei hardware, one of NVIDIA's few true challengers.
DeepSeek just fully opensourced DeepInfer 🔥:
Their in-house inference framework, provides offline inference with the core features needed for serving at scale.
DeepInfer framework supports serving on:
- NVIDIA SM100 GPUs
- Huawei Ascend 950 NPUs
Supported models:
- DeepSeek-V4.1-Flash
Supported features:
- Optimized kernels, model and sampler runtime
- Parallelism: Attention DP / balanced CP, MoE EP + EPLB
- DSpark and confidence-scheduled verification
SWA bounded replay
- Asynchronous scheduling (zero-overhead batch scheduler)
- Batch-invariant inference
- Deterministic sampling
- Structured outputs
Morgan Coder retweeted
💥New paper alert!
All models are equal, but some models are more equal than others.
In our new paper, we investigate whether we can still measure model capabilities after benchmark saturation, by looking at how models solve tasks rather than whether or not they succeed.🧵
Morgan Coder retweeted
It's remarkably short-sighted for websites to block AI agents.
Agents WILL become the way most people transact online - block them and you'll lose a ton of customers.
And I suspect LLMs will get better at browser use faster than humans will improve at detecting them...
notice..
> Delta is blocking AI agents from booking flights
> JPMorgan Chase and Charles Schwab block agents from accessing accounts
> Yelp requires paid data licensing for agents
> Amazon has blocked Grok bot and other shopping agents.
the internet is starting to block AI agents that were supposed to do everything for us
Morgan Coder retweeted
‼️ China proposes banning exports of AI used for cybersecurity.
It wants to ban exports of:
- AI that writes malware beating a mainstream antivirus 85%+ of the time
- AI malware that escapes sandboxes or rewrites 50%+ of its own code per run
- AI agents built for defence or critical infrastructure
- AI models trained on China's "core data"
A licence would be needed for:
- AI that finds 5+ unknown high-severity flaws in a major OS on its own
- AI that builds working exploit chains for major operating systems or browsers
- Weights of models with 2T+ parameters or a global top-10 ranking
- AI chip and memory chip designs
Morgan Coder retweeted
arxiv.org/abs/2606.19535 They backdoored an LLM so that it writes insecure code when runing on specific GPUs!
Morgan Coder retweeted
Is software dead or just not (primarily) built for humans anymore?
I make 4 arguments for each side in my latest essay. A few takeaways:
1. You don’t need to rebuild an entire SaaS to replace it. Many traditional SaaS apps have become incredibly bloated with features. You only have to copy the 20% that matters. I replaced a $200/year YouTube tool after Claude built the one feature that I actually used in 5 minutes.
2. Agents are replacing humans as software’s main users. For example, many people (including me) are getting their agents to create 3D models in Blender without knowing how to use Blender’s UI at all.
3. Building an app is easy, but maintaining it and supporting customers still takes work. Companies want someone to call when things break. I love building my own fitness app, but supporting thousands of other users would turn my hobby into a job.
TL;DR: I think “Software is dead” is BS hyperbole. But I wouldn’t want to bet on software that requires humans to click through screens forever.
The best software companies are thinking agent-first.
📌 Read now: creatoreconomy.so/p/is-softw…
Morgan Coder retweeted
‼️ UPDATE: A researcher says the Broadcom Wi-Fi router bug affects way more countries.
The vulnerable code is also in ZTE's F680 v4, which ISPs from Spain to Vietnam, Egypt and Moldova offer.
Three Movistar Spain routers appear affected. Broadcom still won't assign a CVE.
Morgan Coder retweeted
GH repo here: github.com/ajogalekar/claude…
Morgan Coder retweeted
The computing boom of the 1960s and ’70s revolved around expensive machines from IBM and DEC.
Later decades produced software and internet giants like Microsoft and Google.
I keep wondering whether AI follows a similar path.
NVIDIA supplies the hardware. Model labs supply the intelligence. But the biggest long-term business could belong to whoever turns that infrastructure into a product customers depend on.
If models become cheap and interchangeable, I’d bet on whoever owns the customer and workflow.
Who becomes the Microsoft of this cycle: a model lab, an existing software giant, or a startup nobody’s watching yet?
Morgan Coder retweeted
🚨 BREAKING: A Broadcom Wi-Fi bug lets anyone in range run code on routers without the password.
Quentin Kaiser showed it on Proximus, Orange Belgium and VOO routers and estimates ~1.5M in Belgium are exposed, none patched.
The bug dates to 2010. Broadcom won't assign a CVE.
Morgan Coder retweeted
In the @Research_FRI survey, all categories of experts were similarly concerned about AI risk.
Estimates of the chance of catastrophe by 2100:
Computer scientists: 2-15%
Economists: 2-10%
Industry: 1-9%
Policy: 3-16%
Morgan Coder retweeted
Alexandr Wang's framing on AI warfare cuts straight to the strategic question most people are dancing around.
His thought experiment: put all 5,000 of your AIs on defense. You still lose to an enemy running 10,000.
Scale isn't just an advantage — it determines who gets to run offense at all.
But Wang's real question goes deeper than the numbers: "Can we use our AIs to sabotage their AIs?"
That single question reframes the entire competition. The goal isn't just to build better AI — it's to degrade the enemy's AI faster than they can degrade yours. Offense becomes model poisoning and intelligence ops. Defense becomes system integrity.
Military AI stops being a procurement problem and becomes a doctrine problem. Most governments haven't caught up to that distinction yet.