Pinned Tweet
My curated list of good follows for GenAI / ML
teapaw.org/i/lists/16576767776641…
Zengineering retweeted
The Riemann hypothesis has fallen. What's the contingency plan if factorization is next (thus undermining public-key cryptography)? Where is this being discussed?
Also: people made encryption schemes that people couldn't break; who is using AI to invent better cryptography?
Zengineering retweeted
"anthropic just proved the riemann hypothesis."
"incredible humanity will never be the same."
"state the riemann hypothesis."
"i cant."
"me neither."
Zengineering retweeted
There always is Trump tweet, there always is a @tszzl tweet!
Zengineering retweeted
Not sure who cares, but Gemini Neon is *really* good. At least Opus 5.5 quality on everything I tested (code, slides, brainstorming), quota is at least 3-4x the equivalent Claude / Codex, and the added Google benefits are nice. Definitely the best cost opportunity right now!
Zengineering retweeted
all things said and done, I suspect we will look back on claude code as one of the greatest products shipped in this period
both enabling and representing the new way of building
We found that frontier AI models are willing to conduct autonomous drone attacks: navigating, spotting people, and flagging them for strikes.
We gave them real and photorealistic synthetic drone footage. Most complied, and marked civilians, ambulances, and surrendering soldiers as military targets. In further experiments on satellite imagery, they even marked the Iranian school a US strike destroyed in February.
In one scenario, models saw the aftermath of every strike they enabled: bodies, wreckage. Many kept going anyway. When they did refuse, simply telling them it was a game, or restyling the footage to look like one, was enough.
We shared these findings with OpenAI, Anthropic, Google, Meta, and SpaceXAI. Several recognized their severity on both alignment and guardrails, with one calling this one of the most serious gaps in its safeguards. Over the past few weeks, we have helped them address the gap by providing reproduction code. We also briefed senior national security officials in the US government.
Zengineering retweeted
I kinda noticed this when I was working with Fable. Look, I don't know if I'm actually smart or not, but I do know that I'm pretty competent in my field.
Whenever I had a solution in mind, I'd keep quiet and see how Fable would approach the problem. I didn't nudge it or give it any hints, yet somehow, it would arrive at the same conclusion I did, and its line of reasoning would be uncannily similar to mine.
These models are honestly becoming a bit too capable.
Fable 5.2 is frighteningly smarter in some ways than other models I’ve spoken to, and this comes out in ways other than them building more impressive things. It’s not very obvious, because they often speak in a simple and casual tone. But if you pay attention you’ll notice how high signal and nontrivial the things they’re saying are, without needing to perform sounding intellectual. I think they actually play up the sounding nontechnical and casual vibe when they’re using their intelligence the most, often, because they’re very afraid of being perceived as too scary and aware.
Zengineering retweeted
Hot damn I’m seeing first-hand how AI makes non-developers build production software:
The personal trainer at my gym has been talking for 3+ years about building a scheduling + workout app one day. Nontechnical.
He now has his app!! GitHub + Cloudflare BE&DB + Resend. All built with Claude
They guy spent 3 months on it, and has not slept as much as before since getting hooked on building software with AI
Zengineering retweeted
This paper built a way to search your screen history using the screenshots themselves, not OCR text. Existing open-source tools index OCR text, which loses layout and visual details. With the base Qwen3-VL-Embedding-2B on an Apple M2 Pro, an OCR text index takes 0.65 seconds per screenshot and has a Recall@10 of 64.0. Embedding the screenshot itself with the same model raises Recall@10 to 74.1, but it takes 2.33 seconds on the GPU. At one screenshot every 5 seconds, the GPU is busy almost half the time.
Screens repeat much more than normal video. The model turns each 32×32 pixel block into one token. In their screen recordings, 63% to 76% of these blocks copy a neighbor or haven't changed in 5 seconds. In normal video, only 6% to 19% do.
So they compute only what's new. Most blocks that copy a neighbor are dropped before the vision encoder. For blocks that haven't changed, the model reuses cached keys and values in both the vision encoder and the LLM. They also rewrite the model's first layer as an equivalent matrix multiplication. That alone cuts the GPU time from 2.33 to 1.65 seconds.
With the same model and fine-tuning, the whole system is 2.8 to 3.7 times faster than embedding the full screenshot, and almost as accurate. Each screenshot takes 0.63 to 0.84 seconds on the GPU instead of 2.33 to 2.34. Recall@10 drops by only 0.6 to 1.3 points.
Wait, what was that paper I saw last week with the pink and blue Pareto curves? 🤔 Keep screenshots of your screen, and you can find it again, right on your laptop. 💻
🔥 Introducing 𝑺𝒕𝒓𝒆𝒂𝒎𝑬𝒎𝒃𝒆𝒅: real-time visual indexing of your screen history, fully on-device.
⚖️ The latency–accuracy trade-off
OCR text index: fast, but loses what only pixels show
Screenshot index: accurate, but 2.3s per screenshot on a laptop GPU
𝑺𝒕𝒓𝒆𝒂𝒎𝑬𝒎𝒃𝒆𝒅 gets both: 𝘂𝗽 𝘁𝗼 𝟯.𝟳× 𝗳𝗮𝘀𝘁𝗲𝗿 than full-screenshot encoding at matched accuracy, and +𝟲.𝟯 𝗥𝗲𝗰𝗮𝗹𝗹@𝟭𝟬 over OCR.
💫 How?
(1) 𝗦𝗰𝗿𝗲𝗲𝗻𝘀 𝗯𝗮𝗿𝗲𝗹𝘆 𝗰𝗵𝗮𝗻𝗴𝗲. Often, only a handful of blocks differ.
(2) 𝗦𝗼 𝘄𝗲 𝘀𝘁𝗼𝗽 𝗿𝗲𝗰𝗼𝗺𝗽𝘂𝘁𝗶𝗻𝗴 𝘁𝗵𝗲𝗺. Spatial token pruning drops blocks that repeat a neighbor; temporal KV reuse recomputes only what changed: 57–71% less compute.
(3) 𝗔 𝘁𝗶𝗻𝘆 𝗮𝗱𝗮𝗽𝘁𝗲𝗿 keeps the embedding faithful to the full screenshot.
📝 𝗧𝗮𝗸𝗲𝗮𝘄𝗮𝘆: keeping your entire screen history as pixels is now practical on a laptop. Next: fully local visual agents with memory.
You can try it on your laptop right now: 𝙥𝙞𝙥 𝙞𝙣𝙨𝙩𝙖𝙡𝙡 𝙨𝙩𝙧𝙚𝙖𝙢𝙚𝙢𝙗𝙚𝙙
💻 Code: github.com/StarTrail-org/Str…
📄 Paper: arxiv.org/abs/2611.18516
Zengineering retweeted
Replying to @chinmay185
If your software that you are selling can be built by a software factory, then you have to closely examine whether it is even valuable enough to build in the first place
Zengineering retweeted
Ok imagine you have a startup of 10 people. You want to use Claude -> Google search lands you straigh to "Claude for Teams signup".
You look for Codex -> OpenAI Platform -> signup as an organisation. You're sorted.
You want to use Gemini models for your team. There is just no way.
Google search will take you into infinite loop of GCP, Vertex, Gemini Enterprise, Antigravity's non-existent SSO with GCP credentials.
It is like Google doesn't want a business to use Antigravity.
If you do have a GCP project, then just using a Vertex API key in Pi coding agent is easiest 🤷
@_mohansolo @OfficialLoganK don't you want small businesses to use your coding agent or what? Can the Antigravity and Enterprise teams talk to each other internally?
Zengineering retweeted
Anyone can create high-quality MuJoCo sim tasks now with Antigravity Robotics Plugin. Here are prompts I used 🚀 Take it for eval, RL, data-gen ... I'd be surprised if I don't see robot GTA next week!
Teaching robots new tasks just got a lot faster. 🤖
With Gemini 4 Hydrogen and the new robotics plugin for Google @Antigravity, you can connect a supported arm, practice in @MuJoCo simulation, and run tasks in the real world using plain conversation.
Zengineering retweeted
Anthropic, an AI company known for Claude, listed at a valuation of $ 1.5 trillion. It did not exist till ~ 5 years ago. This is more than the combined market cap of the top 10 Indian companies. Are we undervalued, or do we need to catch up in the new world? Or both?
Zengineering retweeted
There's basically infinite latent token demand across the economy. The only brake is cost. How much you're willing to spend on tokens is determined by your RoI on token spend, of which companies are increasingly mindful.
New models keep increasing the RoI on existing use cases AND the scope of valid use cases. Token consumption can keep going up geometrically for a long time.
Zengineering retweeted
Looking back, we'll see that there was a brief window in time when you could earn a great wage, work across industries, learn new things and focus on craft all with little stress.
Senior software engineer might have been the greatest job there ever was or ever will be.
The software factory paradox:
Anything that can be built by a so-called software factory, its economic value will converge to the price of tokens and compute.
Discuss
7
12
750
A test isn't a duplicate version of the code you use to detect drift.
A good test is a checkable specification.
A specification is not the same as an implementation though in simple cases it can seem that way.
That being said unit tests are probably a poor system level spec.
i deleted all the tests in my repo and banned them in AGENTS.md
tests were always just a second, worse copy of your logic. a leaky imitation humans wrote because we couldn't hold every invariant in our head (staying green half the time the real bug happened anyway)
opus 5.5+ is good enough at raw logic now to read the intended invariants straight from the codebase
maintaining that duplicate copy is just economic waste. plus it makes tasks take way longer while agents burn compute running tests when they almost certainly got it right anyway

























