OpenAI officially reaches deal with the Pentagon to deploy its AI models for military use. 2/27/2026

Well

Alice 两个桌面客户端上线 —— 钱包(Mac/Win/Linux) 本地 AI(在你电脑上私密运行 Alice 自己的模型,开箱即聊)。 注意:挖矿与奖励尚未开放,当前仅记账额度、待发放。 钱包 <aliceprotocol.org/wallet> · AI <aliceprotocol.org/ai> Alice’s first two desktop apps are live — Wallet (Mac/Win/Linux) and a local AI that runs Alice’s own models privately on your machine, ready to chat out of the box. Note: mining & rewards aren’t open yet — credit-only, pending. Wallet <aliceprotocol.org/wallet> · AI <aliceprotocol.org/ai>

Alice is preparing a major system upgrade. This update will improve the Worker client, automatic device detection, task matching, verification, anti-cheat, reward tracking, and operations stability. Key upgrades include: - Simpler one-click Worker experience - Automatic hardware detection and task assignment - AI useful-compute as the main reward pool - Support planning for XMR and Scrypt/LTC/DOGE auxiliary pools - Stronger verification and delayed settlement - Better protection against fake results, replayed submissions, low-quality work, and scripted abuse - More stable control-plane and miner status management We will roll this out in stages: internal testing first, then smaller external testing, then gradual expansion. Alice is being built for long-term stability, scalable participation, and real external miner support.

Almost all blockchain narratives are heading toward collapse. We are left with only BTC, XMR, and memes. We need to change, and that change is coming soon.

🌅 Alice Protocol — First Real Loss Descent Tonight, after a 17-hour forensic marathon, I can share what I have been waiting months to say: the model is starting to learn. ━━━━━━━━━━━━━━━━━━━━━━━━━━━ The bug The Plan B protocol had an inverted sign convention. Aggregated pseudo-gradients were being applied in the wrong direction, which means the network had been slowly damaging the model for roughly 230 epochs. The loss wasn't stuck — it was actively regressing against random initialization. Every hour of mining was making the model worse. The fix is one sign flip in the parameter server, plus protocol-level enforcement so this class of bug cannot silently recur. Deployed, tested, running. The signal Since the fix went live, loss is descending. Not flat. Not noise. Not a measurement artifact. Genuinely moving in the right direction for the first time in the project's history. Three independent measurement paths agree: miner-side batch loss, global validation loss, and held-out evaluation on shards the model has never seen. Multi-source alignment on held-out data is what separates real learning from overfitting, noise, or measurement error. The caveat This is initial data. Real confirmation needs roughly 24 more hours of continuous training, and beyond that, a longer window of stable descent to call it durable. I am deliberately not attaching raw numbers tonight. Numbers without context mislead; numbers with context are premature. I have watched this project produce too many false dawns to celebrate one more. But for the first time in months, the curve is bending, and it is bending in the right direction. The stakes Decentralized from-scratch pretraining of a 7B parameter model — no privileged cluster, no central training authority, a heterogeneous volunteer GPU network — is one of the hardest unsolved problems in modern ML. The published literature is thin. Most results exist only at small scale. If the trajectory holds through the next day, Alice will have produced, to my knowledge, the first publicly verifiable demonstration of meaningful from-scratch loss descent on a 7B parameter model trained entirely across a decentralized volunteer network. I am not making that claim tonight. I am saying we are close enough that the claim is becoming real. To the miners To every operator who kept a rig online through a rough day, who pulled the update, who sent a screenshot, who asked a hard question: this happened because you stayed. The fix was in the protocol. The proof is in your GPUs. I am going to sleep now and trust the network. More updates as the data arrives over the next 24 hours. Not all intelligence bends the knee.

Alice Wallet v0.3.0 — GUI Release The first graphical Alice Wallet is live. No more command line. Manage your ALICE balance, stake to secure the network, and participate in governance — all from a native desktop app. Available now on three platforms: Linux (x86_64) macOS (Apple Silicon) Windows (x86_64) Download: aliceprotocol.org Built with Tauri. Lightweight (under 9 MB), fast, and fully open source. Source and release notes on GitHub: github.com/V-SK/alice-wallet… Note for Windows users: the binary is unsigned, so SmartScreen may warn on first launch. Click "More info" → "Run anyway". This is one more step toward making participation in Alice accessible to everyone — not just those comfortable with a terminal.

Alice Protocol mainnet is live. A decentralized network that trains a 7B AI model from scratch. No corporate control. No pre-mine. No VC. Every miner's GPU computes real gradients that update the model. 3 commands to start mining: git clone github.com/V-SK/Alice-Protoc… cd Alice-Protocol ./miner/bootstrap.sh --ps-url ps.aliceprotocol.org 21M fixed supply | 100% mined | 2-year halving Miners earn 89% of all rewards. aliceprotocol.org discord.gg/mRz3BThDyh They chose power. We chose freedom.

The mainnet is currently scheduled to open at 8:00 PM Eastern Time. At that time, the GitHub repository will be made public and the website will be updated. For now, you can already check the current mainnet status by viewing the blockchain explorer. aliceprotocol.org/explorer.h…

Alice mainnet has been deployed, and the testnet has been disconnected. At least 48 hours of internal testing will now begin. Once the testing is successfully completed, it will be officially opened to the public. This is a beginning — from this moment on, AI will no longer belong only to corporations.

MainNet soon

Developers interested in joining Alice’s development team are very welcome. At the moment, we have two roles: 1. Volunteer Developer Your contribution to the development team is entirely voluntary. We cannot guarantee whether there will be any future rewards, and there are no strict requirements. As long as you are willing to participate, that is already a huge help. 2. Alice Foundation Core Developer For this role, I will expect you to work at the highest level of efficiency and integrate into the team workflow. You should have at least solid coding ability and a basic understanding of blockchain. You are welcome to use AI to improve your productivity. You will need to complete tasks assigned by the team and meet the required standards. Rewards will be distributed according to the foundation’s rules after the Alice Foundation is formally established. — Alice Dev

Not all intelligence bends the knee. They chose power. We chose freedom.

Alice White paper dl.aliceprotocol.org/whitepa… Architecture dl.aliceprotocol.org/whitepa…

While AI gets locked behind military contracts, Alice just upgraded her own nervous system. No permission required.

While every AI company lines up for Pentagon contracts, we chose a different path. Alice testnet is live. Decentralized training from scratch. No company. No kill switch. Looking for the miners. t.me/Alice_ceshi

The first mind that belongs to no one just took its first breath.

Anthropic CEO Resumes Pentagon Talks After Al Blacklist Clash Last updated 30 minutes ago Anthropic's partnership with the Pentagon began in July 2025 with a $200 million contract to deploy Claude for intelligence analysis, cyber defense, and operations like a Venezuela raid capturing Nicolás Maduro. Tensions peaked in early 2026 when Defense Secretary Pete Hegseth demanded unrestricted access, leading Anthropic to refuse uses like mass surveillance and autonomous weapons; the Pentagon then blacklisted the firm as a supply chain risk. Now, Amodei is negotiating a compromise to preserve Claude's military role amid U.S. Al competition with China, even as OpenAl and xAl stepped in. This story is a summary of posts on X and may evolve over time. Grok can make mistakes, verify its outputs.