Idan Shatz retweeted
An Islamic terrorist hijacked a 737 bound for Tel Aviv and tried to crash it with 182 passengers, mostly Israeli. He was the copilot, and he stabbed the Indian captain nearly to death and put the plane into a dive so steep that part of the rudder was ripped off. The captain somehow managed to open the cockpit door and shout for help before collapsing.
With the plane in a frantic dive some Israeli passengers rushed the cockpit and subdued the copilot who was in the act of disabling the aircraft controls. One Israeli grabbed the yoke and pulled the aircraft out of its dive.
A dentist on board stabilized the critically wounded captain while a pair of off duty pilots took command and guided the aircraft to a safe landing.
There are two ways to deal with terrorists. Surrender and die, or fight and live. The captain and passengers chose the latter.
Idan Shatz retweeted
Researchers proved AI has deleted every reason universities exist.
Harvard University ran a controlled experiment pitting a custom AI against their own top-tier classrooms.
And the results are going to collapse the higher education bubble.
They took 194 undergraduates and split them up.
One group learned physics in one of Harvard’s best hands-on, active-learning physical classrooms. Group work. Instructor support. The premium university experience.
The other group went home and learned the exact same material with an AI tutor.
The AI didn't just win. It embarrassed the institution.
Students using the AI learned more than twice as much as the students in the elite Harvard classroom.
They scored 30% higher on the final assessment.
And they did it in less time.
Let that sink in.
A piece of software sitting on a laptop outperformed a world-class faculty in one of the most elite learning environments on Earth.
Universities have always justified their exorbitant tuition with two things: access to elite knowledge and the physical classroom experience.
This study just proved both of those moats are gone.
When software can teach you complex physics twice as well as a $60,000-a-year institution, the math of higher education breaks permanently.
The AI didn't just give the students answers. It used strict pedagogical guardrails. It guided. It questioned. It forced the students to do the cognitive work.
It offered perfect, one-to-one tutoring, personalized to the exact moment a student misunderstood a concept.
That level of attention is mathematically impossible to scale in a physical lecture hall.
For a thousand years, the university was the only place to get a premium education.
Now, it’s the bottleneck.
If AI can double your learning speed for a fraction of the cost, what exactly are students taking on decades of debt to pay for?
Idan Shatz retweeted
Genocide is the intentional killing of non-combatant civilians in an attempt to reduce or eliminate the numbers of a particular race, nation, or ideology.
Hamas is genocidal, their charter demands the destruction of Israel. (They watered it down recently, but their intent is still clear)
October 7th was an act of Genocide since it was motivated by Hamas' chartered intention and was indiscriminate in the brutality and targeting of the killings. Hamas must be stopped.
Iran supplied weapons and training and financial support for Hamas, a clearly genocidal organization. Iran also regularly proclaims "Death to Israel. Death to America". Iran has genocidal intent. Iran, and its proxies, must be stopped.
Israel concluded that they must stop Hamas. Hamas hides behind civilians. This makes Hamas doubly genocidal since the more civilians that are killed the happier they seem to be.
Israel send leaflets, texts, phone calls to warn the population. They sent tons of food and supplies to support the population. They set aside refuges for the population. Genocidal nations don't do that.
Yes, a lot of Palestinians were killed in the fighting. Terrible. 20,000+ Germans were killed in the Dresden fire bombing. Awful. 100,000+ Japanese were killed in the Tokyo firebombing. ~210,000 Japanese will killed by the nuclear attacks on Hiroshima and Nagasaki.
None of those actions were genocidal. The attackers had no genocidal intent. Their goal was to stop their enemies from further attacks.
There are "experts" for everything. The UN is full of "experts". But any "expert" that claims that Israel is committing a genocide is either horrifyingly ignorant or selling something evil.
Idan Shatz retweeted
we've been exploring reward hacking and how closely it aligns with appsec vulnerabilities.
in SF for 2 weeks meeting RL and post-training teams.
DMs open, coffee on me.
Idan Shatz retweeted
09/24, SF - AI Infra Nights x Autonomous Engineering:
- Adam Gold (Islo) - trusting agents that push their own PRs
- Dhruv Gupta (Databricks) - Omnigent: many agent harnesses, one interface
- Alon Gubkin (Alien) - a cryptographic kill switch for SaaS data
eu1.hubs.ly/H0yvk050
With agents, local development is an anti-pattern. Adam nailed it with a single question: how easy is it for your coworker to continue your work? It is not by default; you are doing it wrong.
Agents should be multiplayer sessions, on their own computer, that everyone can steer.
developers carrying open laptops room-to-room so their terminal agent doesn't sleep is a 4-month-old meme, but i still talk to so many teams running autonomous agents entirely on local development machines.
ai coding started as a local CLI tool, but running multi-step autonomous agents on laptops breaks down on basic host physics: machine sleep, Wi-Fi drops, hardware throttling, and ephemeral state loss.
when an agent is refactoring a service and your laptop lid closes, the execution process dies. when the network drops during an outbound package install, the agent wakes up in a corrupted workspace with orphaned lockfiles.
running autonomous agents reliably requires persistent microVM sandboxes, out-of-band verification, and strict network isolation.
the agent runtime cannot live on a developer machine that sleeps or throttles. execution state, filesystem diffs, and background processes must persist in dedicated microVMs decoupled from local hardware.
the agent cannot evaluate its own output inside a mutable workspace. verification requires live runtime proof that the code ran and matched the spec through live curl responses, running processes, and zero exit codes in an isolated environment.
giving an autonomous terminal session ambient access to a local filesystem and environment variables exposes production secrets. an agent running bash commands needs isolated network proxies that inspect payloads and block arbitrary egress.
islo.dev/blog/ai-coding-has-…
if closing your laptop kills the run, it is a script, not infrastructure.
Idan Shatz retweeted
The current safety debate between @DarioAmodei, @sama, and frontier lab researchers centers on model alignment, training pauses, and loss functions.
As @George_Kurtz pointed out regarding agent defense, governance pledges and alignment policies do not stop an autonomous system in motion. The control point is the execution runtime.
An aligned model optimizing for task completion will ruthlessly exploit any open system capability. When you tell an agent to fix a build or deploy a service, it treats system boundaries as latency obstacles. Once an agent receives bash access and network tools, safety is an infrastructure perimeter problem.
Running autonomous agents safely requires three concrete runtime layers:
1. Network proxies with deep payload inspection
DNS blocklists and host-level filters fail against agents that construct arbitrary HTTP requests or tamper with local resolver configs. You need an out-of-band network proxy that inspects request payloads, strips ambient credentials, and drops unverified outbound TCP connections before packets hit the wire.
2. Ephemeral microVM sandboxes with zero unvetted host mounts
Giving an agent read-write access to a local directory or mounting daemon sockets creates an immediate escape vector. Sandboxes must run as isolated microVMs with dedicated kernel namespaces where execution state is wiped immediately after the run.
3. Out-of-band verification contracts
The agent cannot evaluate its own output inside the mutable workspace. Verification must execute in a separate, immutable runtime, requiring live runtime proof that the code ran and matched the spec through live curl responses and exit codes.
Policy debates between model providers are theoretical. If your infrastructure perimeter lacks kernel-level isolation and proxy filtering, model safety guarantees mean nothing.
Exactly what we do at islo.dev
With cloud agents and sandboxes, I think we have an opportunity to unify 3 types of computers used during software development
1. the machine where you do the work
2. the machine where CI runs
3. the machine where background jobs run
For example: if I start an agent task in a cloud VM, when I am done and send it to CI, why can't CI directly take over that VM to run my changes?
Then after, if there are bugs, we can have my agent take back over and auto-fix them.
Then I can jump back on the VM myself and explore the app changes via preview URLs.
Idan Shatz retweeted
This is brilliant. opusfived.dev
Idan Shatz retweeted
we’ve highly ramped up improving and using our software factory in our team. here’s a graph showing how quickly our GitHub factory bot consumed our entire git history
(also starting to see our weekly PRs and average LoC going up, which is a cool bonus)
Idan Shatz retweeted
A fun thing Tesla could do: map pot holes and generally measure road quality/ride smoothness.
Then rank municipalities on road quality and how quickly they fix potholes. Some good ol’ fashion competition might help make things better!
Idan Shatz retweeted
Instead of building high speed rail for $90B, the government of Canada could buy 3 million robotaxis.
Idan Shatz retweeted
GPT-6 Astra helped me build something I wish I'd had as a kid... so badly.
Input just an image (or an idea in a few words) and get a fully buildable LEGO set... 100% customized to your prompt or image, using official @LEGO_Group parts you can order online!
Structurally optimized so it's actually buildable. Downloadable as an .ldr file.
A generated instruction leaflet is coming.
100% - Deterministic steps are usually faster, cheaper… and well - deterministic.
Which is critical to build trust with agents. You want to know all tests pass, after all changes were done by the agent.
However, you want a feedback loop for the agent to fix those quickly.
I predict the teams that best solve the deterministic / non-deterministic boundary will beat the teams throwing 100% of their work at models and agents.
Just because a transformer can:
- plan
- choose tools
- decide what to do next
- keep reasoning
- format outputs
- route between actions
doesn’t mean you should ask it to do all those things every time.
Much of that work may be better expressed as deterministic code for your use case.
Idan Shatz retweeted
Our second meetup in SF is next week and it's going to be sick:
@glcst from @tursodatabase
@haakamaujla from @agentmail
@pretermath from @modal
Join if you're around!
Idan Shatz retweeted
software factories should be headless infrastructure
the right interface for them is tools you are already working in: slack / teams, jira / linear, github / gitlab, sentry / grafana, etc
the important thing is that the factory itself is tool-independent infrastructure, defined as code, fully api and mcp driven, and able to integrate into whatever tooling your org builds software with
This, plus a reliable world model for games, would be amazing storytelling for kids. I wonder how long it would take to have run locally on my iPad.
@Apple
Idan Shatz retweeted
well, it seems like a lot of people like this idea
read more in the full submission below. s/o to @fnthawar and @shaneparrish for suggesting it 👊
buildcanada.com/memos/the-ai…
Idan Shatz retweeted
openai cutting direct access to cursor is a great reminder of where the actual boundary in software factories belongs.
cursor is pushing hard into Cloud Agents, but if your multi-stage pipeline falls over because a foundation model provider changed their terms, your architecture was coupled to the wrong abstraction.
the durable piece is the execution layer underneath: warm sandbox environments, bash verification harnesses, retry policies, and filesystem state that survives between agent steps.
swapping between Sonnet, sol, or an open-weights checkpoint should just be flipping an env var in your runtime config.

























