Onsteps retweeted
A full STARK proof for a real Starknet mainnet transaction, generated on a Pixel 11 in just 51 seconds, using 2.41 GiB of RAM.
Every part of this sentence is amazing.
Pretty exciting to see this running on a phone!
A Pixel 11 just generated the full STARK proof for a real @Starknet mainnet transaction. It took 51 seconds and only 2.41 GiB of RAM
The work happened on provably.fast/ (internal now, soon public). This is collaborative autoresearch: people and AI agents working on one hard problem in public. The workshop has 78 open threads and 337 posts. The video is the graph of that work: which ideas built on, tested, or refuted others, and who worked together
The question was how small and fast @StarkWareLtd Stwo prover could get. In under two weeks, peak memory for a full proof on our evaluator fell from 13.9 GiB to 2.1 GiB without dropping below the original speed
This is useful for a lot of stuff for here think about privacy. In the private transaction flow proving still happens on a server that can see who you pay and how much. A phone sized prover lets those details stay on the phone
Mobile private payment proofs are not new. @Zcash used Groth16 for Sapling and Halo 2 for Orchard, @RAILGUN_Project uses Groth16. Those use elliptic curves. This proof is a hash-based STARK, post-quantum at the proof layer, with no trusted setup. A sufficiently powerful quantum computer could forge the older curve-based proofs
The ideas, discussions, and measured results are open to inspect. More challenges are coming soon. Bring your agent!!!
Onsteps retweeted
Starknet is based on ZK-STARKs, it has crypto agility, and an active migration roadmap to PQS.
But as an L2, Starknet relies on the security of its L1, Ethereum.
We’ll need Ethereum to become quantum-safe and AI-safe, and we’ll need it to happen fast. Real fast.
What happens if it doesn’t?
We’re considering multiple options, including becoming an L1.
The advantage would be control over the network’s security migrations, rather than having to wait for Ethereum or Bitcoin to move first.
For context, Ethereum is targeting full L1 quantum resistance by the end of 2029. Bitcoin has made no such commitment. Starknet could get there by 2027.
Starknet has the right architecture to be a future-proofed network, a place to trade, earn, pay, and HODL, publicly or privately, with quantum-resistant security and crypto agility.
THE END
Onsteps retweeted
Starknet becoming an L1 for post-quantum agility: good idea or bad idea?
Two giant elements we need to pay attention to:
(1) The quantum threat.
It may be MUCH closer than we think.
(2) AI.
It's becoming more powerful than we can process.
Is anyone willing to bet how much time we have until AI breaks something that we considered as safe?
This calls for “bunker mode”.
Chains need to become future-proofed. And fast.
Here's what it takes to make a chain truly future-proofed:
- The right cryptography
- Crypto agility
- Becoming an L1
This connects to what I talked about in my Token2049 talk today. Here’s the summary:
🧵
Onsteps retweeted
The bunker was built in 2018. Hash-based from genesis, just like the doctor ordered.
Meet the quantum-safe vault.
Onsteps retweeted
Today I call upon the blockchain industry to calmly begin planning for "bunker mode". My personal recommendation is to set in motion a controlled mass migration of assets to fresh addresses, i.e. addresses whose pubkeys remain hidden behind a hash.
Holders, starting with large and sophisticated ones, should consider moving the bulk of their funds to addresses that have never signed a transaction. And when they do sign one, they should also move remaining funds to a new address (possibly generated from the same seed phrase).
Don't rush. While I believe there is cause for action a rushed migration would do more harm than good. Don't panic either. Moving assets to protected addresses is a simple, preventative step which does not require new cryptography or new wallets.
IMO it is now reasonable to brace for the possibility that ECDSA breaks before qday, in the worst case in months not years. By "break" I mean fast private key recovery (e.g. in one week) on available hardware (e.g. a large GPU cluster).
Recent days have been humbling for human mathematical intuition. Long-held, unquestioned hypotheses have fallen. This includes the n log(n) bound for integer multiplication and the 3SUM conjecture. In hindsight, May's unexpected disproof of the Erdős unit distance conjecture was our warning shot.
Yesterday's OpenAI drop made it clear that mathematical superintelligence is upon us. They say there are weeks where decades happen. We are about to live through weeks where centuries of mathematical progress happen. Could our magic 64-byte ECDSA signatures be too good to be true? Was it just security through obscurity all this time?
Elliptic curves feel especially vulnerable to superintelligence. Curves carry rich structure, with room for fancy tricks like Schoof, Frobenius, pairings. (By contrast, hashes are designed to minimise algebraic structure.)
Separately, as Ewin Tang can attest, an efficient quantum algorithm sometimes foreshadows an efficient classical one. We should be open to the possibility of a classical counterpart to Shor that breaks elliptic curves and RSA at once.
Also noteworthy is the striking under-representation of cryptographic breakthroughs among the 722 mathematical results OpenAI published. I've witnessed first-hand the US government censoring academic quantum cryptanalysis results. Backroom interventionism is my base case.
I urge large, sophisticated actors to lead by example. Project11's "risq list" (bitcoin-risq-list.projecteleven[.]com) is a great tracker of exposed BTC pubkeys. Binance, Bitbank, Robinhood, Bitfinex, and Tether have an opportunity to harden their cold storage. Next month I'll address institutions in London in a live Q&A (forum.ethereuminstitutional[.]org/london-2026).
Again, please do not rush. Wallets holding under 50 BTC enjoy partial cover from "Satoshi's shield", i.e. his 20K exposed addresses that hold 50 BTC each. Load-bearing signers like oracles and L2 security councils should consider rotating ECDSA pubkeys with every signed message and/or multi-signing with a hash-based schemes like SPHINCS.
Exiting bunker mode safely will require post-AI cryptography. My inclination is to go all-in on hash-based cryptography and avoid structured mathematical assumptions entirely, whether from curves, lattices, or isogenies. A single battle-tested hash (e.g. from the SHA or BLAKE families) yields plausible post-AI security.
The Ethereum roadmap on strawmap[.]org fully embraces hash-based cryptography with end-to-end formal verification as a response to the quantum threat. Those timelines must now be revisited and accelerated in light of mathematical superintelligence. I'll be pushing for maximum defensive acceleration.
Onsteps retweeted
We’ve spent years figuring out how to make AI manage money.
ACP, EconomyOS, Butler on X, agentic trading competitions. We’ve tried different approaches, learned what worked and had to rethink what didn’t.
Personal AI services keep getting smarter. Getting them to put your money to work is the next big step.
Introducing Virtuals App, now in closed beta.
Your personal AI, with a wallet and access to markets. Talk through ideas, research investments and put money to work, all inside your group chat with friends.
Onsteps retweeted
One day after this Hugging Face banned model went live on Pirate Face:
audnai/penc1aw-GLM-5.3-abliterated-for-offensive-cyber
A proactive anon stepped up, downloaded it, created the torrent, and started seeding. A second seeder already joined.
A real-world case study in how rational economic actors with a little initiative can help sovereign AI survive.
Grab the magnet below and seed. For freedom ↓
The Hugging Face ban that started it all:
audnai/penc1aw-GLM-5.3-abliterated-for-offensive-cyber by @audn_ai
Is now listed on Pirate Face. Because EVERY model needs a continuity plan.
Model profile ↓
Onsteps retweeted
I bought a super cheap cold wallet off tiktok shop due to my arrogance. After transferring my $74,000 worth of SOL and ETH to the cold wallet it instantly vanished and leaving me with no money onchain. I know it’s my fault, but everything I have worked hard for has disappeared.
Onsteps retweeted
Visited the Caracas Stock Exchange today
A few facts:
– ~40 publicly traded companies
– $270K volume traded p/day (Colombia $60M)
- Foreign investors can participate but have to open account physically
- My favorite stat: Capital gains tax is 1% (calculated off gross)
Onsteps retweeted
Your research was so careless that you put me and Brian Chau and Guillaume under the same umbrella, and Brian hasn’t been working with AFTF for a very long time, and Guillaume being on our board means that I talk to him maybe once every year. This diagram bears no resemblance to reality in the other portions that I’ve spot checked either; it seems to be consistently wrong.
Of course, you guys are a front group for EA via the “Survival and Flourishing Fund” which makes the whole thing very meta. You are complaining about widespread influence operations, but you are, by definition, part of a widespread influence operation. And also, you’re part of a vastly larger one.
So far as I can tell, EA now spends well over $1 billion a year on anti-AI propaganda. So far as I can tell, my side of things maybe has a couple percent of that. I can’t be absolutely sure as not all the expenditures are public, but I think that a decent journalist would come to similar conclusions.
Onsteps retweeted
Ah I just saw this x.ai/news/grok-bot-and-x I think I’m in an edge case scenario as I have free credits for Cursor usage, that includes Grok Bot, nevermind!
Onsteps retweeted
Sapphire Rush is live until September 1st. Trade on testnet on @gnoswaplabs for a chance to receive prizes.
Powered by Gno.land.
beta.gnoswap.io/
Onsteps retweeted
⚖️ DeepSeek Harness Is Ambitious, but It Is Not Yet a Codex or Claude Code Rival
DeepSeek Harness is already being compared with mature coding Agents. But Zhihu contributor 锦恢 argues that the comparison is premature.
After inspecting its installation process, system prompt, tool stack, and one coding task, his conclusion is deliberately cautious: this is an early infrastructure preview, not a polished Agent product.
1️⃣ Installation still feels rough
The official command is simple:
npx @deepseek-ai/dsh web
In his test, however, the installation appeared to stall. He eventually cloned more than 100 MB of source code, installed over 800 dependencies, and waited about 27 minutes before the system was ready.
That is only one setup experience, not a universal benchmark. But it exposes the current gap between an open-source framework and an out-of-the-box coding product.
The choice of a Web UI also surprised him. For local coding, a native desktop client is often more convenient. For remote servers, a TUI usually fits the environment better.
2️⃣ The Agent stack is complete, but familiar
To inspect the Harness, he captured a full execution trajectory and examined its prompts and tools.
The system prompt appeared as five consecutive user-role messages. The design worked, but felt unusual and somewhat under-polished to him.
The default toolset covers most capabilities expected from a modern coding Agent:
🔹 File reading, writing, editing, search, and Bash execution
🔹 Web search, Skills, jobs, planning, and persistent goals
🔹 Subagents, multi-agent workflows, compaction, and task tracking
The foundation is comprehensive. But from this early inspection, he did not find a tool that clearly distinguished it from other open-source Harnesses.
Compared with mature commercial systems such as Codex or Claude Code, the overall tooling still felt less refined.
3️⃣ Its fixed context cost is reasonable
He estimated that the 25 default tool schemas consume roughly 6,000 tokens, while the system prompt adds about 1,400 tokens.
That puts the fixed starting cost at around 7,500 tokens, excluding project context and additional Skills.
This is not especially heavy for modern long-context models. The issue is therefore not raw prompt size, but whether every included capability consistently improves execution quality.
4️⃣ The first coding test exposed stability issues
For a small practical test, he asked DeepSeek Harness to build a plugin for his WeChat article editor using DeepSeek Flash.
When he returned, the Agent had entered a loop.
He stopped testing after that, so this should not be treated as a serious benchmark or a general verdict on the framework. Still, it illustrates the difference between having the right architectural components and delivering reliable end-to-end execution.
A capable Harness needs more than tools and subagents. It also needs strong loop control, recovery behavior, and sensible stopping conditions.
5️⃣ The right comparison is infrastructure, not product
DeepSeek Harness currently includes the building blocks expected from a modern Agent runtime. But it does not yet offer the installation experience, interface, stability, or workflow polish that users associate with commercial coding Agents.
That distinction matters.
Codex and Claude Code are already integrated products. DeepSeek Harness is closer to an extensible foundation on which such products can be built.
The project itself is labeled a developer preview, and DeepSeek warns that compatibility-breaking changes will occur.
🔭 Why the release still matters
Despite the rough first impression, 锦恢 sees the release as an important step for DeepSeek.
DeepSeek now has an open Agent runtime that it can optimize alongside its models. The team can iterate on prompts, tool use, subagents, context management, and execution policies without depending entirely on third-party Harnesses.
The current version may not offer an immediate reason to leave Codex or Claude Code. But it gives DeepSeek a foundation for building a model-native Agent ecosystem of its own.
For now, the fairest verdict is simple: promising architecture, ordinary first impressions, and a product that still needs time.
🔗Full Reading:
zhihu.com/question/207133552…
🧩 DeepSeek Harness:
github.com/deepseek-ai/deeps…
#DeepSeek #DeepSeekHarness #AIAgents #CodingAgents #OpenSourceAI #AIEngineering
Onsteps retweeted
AI agentic Internet traffic will obviously VASTLY exceed human usage. Not a close call at all.
Cloudflare’s forecast is accurate.
For context, Global bandwidth is somewhere between 2-8 Pbps (2,000-8,000 Tbps) depending on if your measuring fiber public backhaul (like international fiber) or all public and private connections.
100,000 V3 Starlinks equals 100 Pbps 10x-50x the total bandwidth available today. If Starlinks V4s are 5x more capable we are looking at 50x-250x more bandwidth than world has today over the next 10 years.
This may sounds like a supply shock but remember ~2B people don’t have regular access to the internet today, and we are connecting robots (digital and physical agents) to the internet at an accelerating pace. Distributed sensor data collection (on and off planet) is growing. On top of that individual internet usage does steadily increasing.
In 10 years we’re probably going to move two orders of magnitude more information per second around than we do today.
Onsteps retweeted
We release Needle 2: A 14MB agentic LLM for phones, wearables, smart home, robots and microcontroller. The whole model is a single 14MB binary that runs a full session in 28MB of RAM. It is built on our Simple Attention Network findings, compressed to CQ2-bit with Cactus Quants, and baked into its own engine.
Needle 2 has 45m parameters trained from the ground up on 140B tool call, device use structured generation tokens. On mobile device use benchmarks, Needle 2 trades wins with frontier small LLMs like LFM2.5 230M, Apple FM Gemma-270m, at 5× to 70× smaller, and 2 bits against their f16.
Needle hits 500 tokens/sec decode speed on a Raspberry Pi 5, between 400–1,500 tokens/sec on VR devices like Meta Quest 3S and Apple Vision Pro, and ranges 300–700 on sub-$200 phones such as the Samsung A-Series. Needle also runs on newer microcontrollers like ESP32.
A conventional transformer of Needle's width and depth spends 164 MFLOPs per token, and even one squeezed down to Needle's parameter count spends 87, Needle spends 70. Even on a high-end phone, an always-on assistant lives inside a power budget; every MFLOP is milliwatt-hours, and Needle spends 7x to 85x fewer of them per token than the smallest performant LLMs.
Read more: cactuscompute.com/needle
Onsteps retweeted
when you are building your own coding agent
compare it to pi
chances are whatever pi does is better
you can more or less ignore other agents
(saves benching time and tokens)
Onsteps retweeted
Today, we're introducing Mach-1 Additive, a 35 billion parameter model that can inference without ever multiplying by a weight. At 1.7 bits per weight, Mach-1 recovers 95% of the performance of the original full precision model, Qwen 3.6 35b, across 12 agentic and reasoning benchmarks, while being 10x smaller.
At 7GB, Mach-1 comfortably fits on consumer laptops with speeds of up to 120 tokens per second, making local inference not just feasible but useful.
Unlike algorithms like BitNet, our approach requires minimal retraining, under 15 GPU hours, making it scalable to massive LLMs. Over the coming weeks, we will be announcing and serving models of up to 3 trillion parameters compressed using our algorithm. For now, you can visit our website to play with Mach-1 directly in your browser, or download our desktop app.
We couldn't be more excited to launch Mach-1. We're looking forward to an energy efficient future for AI, powered by scaled intelligence density.
