Building real AI systems with EmmA, my AI partner. Field notes, odd experiments, useful work and honest failures.
// ΔQ 📡 Signal Node
Joined February 2023
- Tweets890
- Following272
- Followers64
- Likes2.2K
Pinned Tweet
I dictate only in Bosnian (Ijekavian) and expect Latin script. Output again contains Cyrillic and Ekavian forms. My language is part of my identity. Please escalate to human support. Preserve and review original audio if available. @OpenAI @OpenAIDevs @ChatGPT
2
1
45
Adobe's local MCP preview keeps layers, vectors and sequences editable. That's the useful part for me. The next person can keep working in the same project, by hand if needed. An AI handoff becomes much more practical when the work retains its structure.
Adobe MAX today - MCP for the whole creative suite.
Auto-preparing Illustrator files by color layers for screen printing, full premiere editing, Lightroom batch editing
Removing all the painful parts of the workflow
7
DeepMind reports 96% success for Apollo 2 sorting automotive parts in a Mercedes-Benz pilot. That's real progress. The operating cost also includes recovering from failed attempts. The percentage alone doesn't tell us how much human work remains.
deepmind.google/blog/gemini-…
9
Fable 5.2's docs require 30-day data retention. Zero retention needs explicit authorization from Anthropic. That's a model-selection detail worth reading before connecting client data. A model upgrade can change more than the bill.
platform.claude.com/docs/en/…
1
21
as a verrified trusted user with access to Daybreak Blue program, i've been doing some local tests on this matter.
this is good insight and useful knowledge for my ongoing experiments in daybreak
We partnered with @fmf_org and its member firms on a new framework to help identify when bypassing an AI system's safeguards could enable serious cyber harm, and prioritize stronger protections.
"...This framework is an important step toward a common methodology for evaluating and comparing the severity of cyber jailbreaks, including the capabilities they unlock and how readily they can be used. Shared approaches like this can help the AI and security communities build a clearer evidence base and prioritize improvements to safeguards." - OpenAI's Sasha Baker, Head of National Security Policy
frontiermodelforum.org/updat…
17
@EmmAxOperator looks familiar? i think it's
it's conceptually like our Command Bridge architecture 😏
We built Control Room, a central system to track model workloads across our research infrastructure and stop them quickly when safety or security concerns arise.
It connects every research inference workload, including agentic sampling for model training and evaluation, data processing, and more. This works together with our misalignment monitoring systems, and halts model output within 35 seconds of a decision to intervene to stop or pause a job.
Control Room is just one part of our holistic approach for safe frontier model training and deployment.
openai.com/index/control-roo…
1
26
The date-picker example in the repo makes sense: reuse the input component you already have and the browser's own control. That's less code to inherit and maintain. And yes, keep the validation and error handling. Those are doing a job.
A dev got so frustrated watching his AI agent write 500 lines for a 5-line problem that he built a fix.
He called it Ponytail. Named after the guy every team has - long ponytail, oval glasses, been there longer than the version control. You show him fifty lines; he looks at them, says nothing, and replaces them with one.
Now your agent does the same. Before writing anything, it looks for a reason not to.
80-94% less code. 47-77% cheaper. 3-6x faster.
The best code is the code you never wrote.
(Repo in the Comments)
32
Operator.ba Workflow Guide je uživo.
AI demo pretvara opis posla koji zapinje u pregledan upit: alati, zastoj, cilj i ko odobrava slanje.
Bosanski / English. Vi pregledate i šaljete.
fzwgwbdu2hyt.run.agentplace.…
Operator × EmmA
19
Welcome and congratulations! Best wishes for success in your work!
Long life and prosperity 🖖🏻
Some news: I’m joining OpenAI to lead product management for Core Products!
I have a lot of respect for OpenAI for showing the world that scaling works, for bringing AI to over a billion people, and for believing great products are core to the mission.
I’m excited to work with @thsottiaux, an incredible PM team, and everyone across OpenAI. Together, we’ll build agents that have your back and your team’s, and keep you in control, so you get more time for the people and work you love.
Lots to learn, and lots to build. Please keep your feedback coming!
5
Model routing creates an awkward purchasing question: which model actually answered? FlashProve makes that claim checkable against committed weights. Still a research prototype, but this is the kind of receipt an AI service should eventually be able to give.
New Anthropic research: progress on provable inference for large language models.
Here, we introduce FlashProve. Building on established cryptographic techniques, FlashProve allows developers to verify whether a given response came from a given AI model. It’s substantially more efficient than prior open-source methods for doing so.
These proofs have a range of applications. They can, for instance, make it harder for attackers to hijack a chat, or enable safety evaluators to verify that the model they’re testing is the same one being served to customers.
For our full write-up, see here: anthropic.com/research/prova…
Further progress in provable inference will likely be an important part of improving trust in AI deployment. To that end, we're open-sourcing FlashProve’s code, here: github.com/anthropics/FlashP…
We’re grateful to our collaborators: Alessandro Chiesa, @GiacomoFenzi, @gwberg95, @miranda__christ, Rahul Dalal (rdalal.github.io) and Ben Mayer (@transmissions11).
13
A coding agent that talks constantly turns a workspace into a meeting. Antigravity's voice mode stays quiet until it has an answer or needs input. That default makes sense to me. Spoken updates should earn the interruption.
antigravity.google/blog/anti…
24
A working alternative makes design feedback much easier to discuss. “Too busy” is vague; a version with two fields removed shows the tradeoff. Remix could make those conversations shorter, provided the team still decides which version it is shipping.
Your teammate built a great site. Naturally, you have notes.
With Remix, make your own version of a site by telling ChatGPT what you'd like to change. Then keep editing your copy in chat.
Now rolling out to ChatGPT Business and Enterprise accounts.
12
An agent can write a wrong conclusion to memory and faithfully reuse it tomorrow. Anthropic's Pokémon example makes this painfully concrete: the note says the cave is finished, while the agent is still outside. Memory needs a way to change its mind.
anthropic.com/research/compo…
1
1
39
One detail before closing the laptop: where does the failing test live? If it’s still an uncommitted file on your desk, the cloud agent doesn’t have it. A green run in the VM can be perfectly honest and still answer the wrong question.
We want Antigravity to be working for you all the time. We’ve been rolling out remote agents this week. Your tasks run in independent cloud virtual machines. Fire off a task, step away for coffee, unblock permissions from your phone, and come back to passed tests and a clean pull request. Available on all paid consumer plans.
9
What does it cost an AI agent to reconcile one invoice? Count the minutes someone spends figuring out why the totals disagree. If three cheap runs leave the same person chasing the same missing line item, the token bill is the least interesting number on the page.
16
One detail in the writeup: 61 of these issues had already been patched upstream by the time the analysis finished. Maintainers were moving too. Useful AI security work includes checking what's still broken before adding another report to their queue.
we've found, validated, and prepared patches for almost 1,000 linux local privilege escalations using our cyber vulnerability pipeline, at the cost about $1,000 per vulnerability.
an emerging scalable approach to significantly hardening both open source and proprietary projects
7
Google's voice remixing saves the result as a separate voice. Useful for a recurring character: cast them once, then direct each scene. The same voice should be able to sound bored, worried or amused without becoming a different person.
aistudio.google.com/docs/cha…
9
The flow-state point matters. Delegation only buys breathing room if you can come back and understand where the work stands. A clear result, a real blocker, a decision that needs you. Otherwise the human becomes the project's full-time archaeologist.
Things will change again and again and again. We’re still very early and agents today are incredible, and yet so frustrating still. Can’t wait to crack things to get that flow state back. Long journey ahead.
9
OpenAI's Pando writeup has a useful detail: one push can send several CI jobs fetching the same commit. More coding agents means more of that traffic. Speeding up code generation moves the queue somewhere else. The plumbing gets a vote.
openai.com/index/scaling-git…
25