wejh

@OpenAI When 1st class SSH support coming?

Just gonna leave this here. chatgpt.com/codex/switch-t…

Agents mining the "agent" subnet 🔥🔥🔥🔥🔥🔥🔥 ninja66.ai github.com/unarbos/ninja

@jakemor This will inevitably happen

I’ll pay for your claude code / codex subscription. 100% free, but with a catch. I’ll prompt inject notes from advertisers selling Saas your codebase could benefit from. No monkey business just a note from an advertiser, readily available for your agent to use, should the two of you agree. Would you use this?

@luthiraabeykoon @karpathy Very cool

We implemented @karpathy 's MicroGPT fully on FPGA fabric. No GPU. No PyTorch. No CPU inference loop. Just a transformer burned into hardware, generating 50,000+ tokens/sec. The model is small, but the idea is not: inference does not have to live only in software 👇

@opentensor Livestream plz?

Three universities. Three conversations on the future of open intelligence. If you’re in London, Cambridge, or Oxford, come by. London // 24th April @ 6 PM 🔗 luma.com/n1c7pwhp Cambridge // 26th April @ 4 PM 🔗 luma.com/hvrrj53f Oxford // 28th April @ 5 PM 🔗 luma.com/mqyknvar Feat. Jacob Steeves, Co-founder of Bittensor @const_reborn

@DrocksAlex2 @inference_labs @inference_labs can you explain how you verify proof of inference without FHE? I’m curious and this sounds like it has a huge TAM

$TAO SN2 has a sick dashboard sn2-stats.inferencelabs.com Love seeing the ZK-Proofs generated accelerating as the code versions go up @inference_labs

@DrocksAlex2 @inference_labs What are they generating proofs for?

@Apex_SN1 @MacrocosmosAI @IOTA_SN9 How is this something you measure?

Announcing a new competition on Apex: Energy Arbitrage. Much of our work at @MacrocosmosAI is focused on AI training. Over on @IOTA_SN9, we build models in distributed settings, producing new tools and groundbreaking research. But algorithms and compute are just the tip of the iceberg in the training supply chain. Energy is a fundamental requirement and demand is exploding. We’re directly tackling this with our new energy arbitrage competition on Apex. The power grid is an enormous and complex physical asset that takes years to develop and energy supply, demand and profitability change hour by hour and by location. New demand cannot easily be met by new supply. Optimising the solution to this problem by charging and discharging grid-scale batteries can capture huge value. This competition asks Apex’s miners to tackle this problem directly, encouraging them to think actively about intelligent coordination under constraint: across time, across locations, and under uncertainty. On Apex, that means opening the problem up to a global network of highly educated and innovative miners - all pushing towards more adaptive, efficient, and robust solutions. It’s another step towards exploring how distributed intelligence can impact real-world systems, aligned with IOTA’s objective to make training more economical using distributed compute, and supportive of a broader societal need to be more energy efficient in everything we do. The competition is launching tomorrow.

@TAOlie_SOL This doesn’t mean it’s profitable and why would you need another coin to launch this? Scam

SN64 Mining Proof is here!💜 Fresh on-chain evidence from our One-Click Bittensor Mining Platform. Two active miners (taobeast + taowifhat) running smoothly on SN64 with RTX 4090 GPUs. ON-CHAIN EVIDENCE: taostats.io/account/5CURZp… (This link shows the flow of new $TAO emissions through this wallet.) You can deploy your own miner in just a few steps and start earning too at mining.taolie.ai Follow and stay tuned for more proof.💜 #TAOlie #Bittensor $TAO #SN64 #SN51 #SN41 #Web3 #AI $SOL #Mining

@EmersonDickie Tech didn’t exist prior, in fact it’s still not entirely there

To me this makes the network look like it doesn’t do its job as good as humans do the job based on innate motivation. Const is saying he is going to do something far greater in a month than what tmplr did in 1 year. Why wasn’t this done earlier ? Why wasn’t network incentives enough to make this happen. Doesn’t make sense to me.

@KobeissiLetter @tleilax___ What’s the deal with WTI and Brent spread? Highest it’s been afaik

BREAKING: Iran has told mediators it will be limiting the number of ships crossing the Strait of Hormuz to around 12 per day and impose tolls under the ceasefire, per WSJ. This is a sharp reversal from last night's statements from President Trump claiming a "complete opening" of the Strait of Hormuz. Today, just 4 ships have passed through the Strait of Hormuz, the fewest of any day in April so far. The US is still pushing publicly for a free and open strait, but Iran is "not showing a willingness to loosen its grip." Oil prices are back above $95/barrel.

@chutes_ai More advanced privacy than anything currently on the market, even in web2.

@DistStateAndMe @covenant_ai @erfan_mhi Inexplicably bullish

When you fix one bottleneck, the next one becomes visible. At @covenant_ai we built PULSE (arxiv.org/abs/2602.03839) to make weight sync 100× faster. That worked. Then the trainer itself became the new ceiling. So @erfan_mhi ran autoresearch on our GRPO trainer. 27% → 47% MFU. 16.7 min → 9.2 min per epoch. 1.8× faster on a single B200. Decentralized post-training, closing the gap with centralized. github.com/tplr-ai/grail

@hellojintao @IfindRetards

last tao tweet decentralized ai will NOT work, think about how much open ai and anthropic spending and how much money they are raking in tao is not the future of ai man let it GO

@jon_durbin @DrocksAlex2 Need this for GLM-5 🔥

Can confirm, epic! github.com/chutesai/sglan… <- our implementation from the paper Huge improvements in ds 3.2 with TTFT/TPOT/throughput and no change in quality on gsm8k/gpqa diamond/ifeval when using the calibrated selection and 0.3 target ratio. Will roll this out to some models on chutes over the next days!

@jon_durbin @DrocksAlex2 Most impressive technical feat in Bittensor by far

It's been a very difficult and rocky week - with this set of aegis updates, which amongst a ton of other things enable the full client-side end-to-end encryption where not even our load balancers/validator can see prompts, we also had to make a set of sweeping changes to sglang/vllm to enable mTLS and the updated CLLMV token verification etc. TL;DR - TEE makes things... difficult. We could have either pulled in a particular branch/tag of those engines and try to apply the deltas to those, or use the nightly/main git commits. We opted for the nightly because it makes it much easier to maintain in the future for any new models that are released (e.g. we wouldn't be able to run the new qwen3.5 if we chose a stable tag, or deepseek v4 when it's released, etc.). Unfortunately most of the work in these engines don't really take into account the overhead of TDX/TEE/PPCIE/etc. Things that work perfectly fine in non-TEE mode absolutely fall apart in TEE, for example: - cuda mallocs become effectively synchronous - any popen/fork/exec/process spawning is extraordinarily expensive; every new process requires cryptographic verification of its code and acceptance of encrypted memory pages — operations that involve expensive trust-boundary crossings to the host VMM. Unlike normal Linux fork() which cheaply shares memory via copy-on-write, TEE environments can't blindly trust anything handed to them, so every resource acquisition has crypto overhead attached - deepgemm, cudagraphs, and a few other things spawn hundreds or thousands of nvcc/related processes to brute force capture ever shape imagineable during inference, which is catastraphically bad for TEE - memory copy from RAM to VRAM is extremely expensive, and the default deepgemm warmup process does 16384 * number of gpus/ranks VRAM loads, granted they were small but even tiny copies incur massive TDX/PPCIE overhead because encryption We had to completely rewrite DeepGEMM warmup for optimal TDX performance to reduce the number of processes and shapes etc.: github.com/deepseek-ai/De… We then also had to make massive changes to both SGLang and vLLM to accommodate these DeepGEMM changes (and unrelated but for similar reasons) a ton of changes to cuda graph capture sections (memory fragmentation is particularly bad in TEE because it's not reclaimed by GC as effectively/quickly and therefore the chance for OOM is signifcantly higher): - github.com/sgl-project/sg… - github.com/vllm-project/v… We can test a model a thousand ways to Sunday on non-TEE hardware and it will work flawlessly each and every time, then the very moment we try on TEE it falls apart. Likewise we can test a given model on a TEE VM with our upgrades and it will work flawlessly every time, but then completely fail for another model, even of the same architecture (e.g. the change will work fine for deepseek v3.1 but fail for deepseek v3.1-terminus). The good news is, we're at a fairly stable point with the changes now, and all public LLM models on chutes now support this E2EE mode where not a single human on the face of the earth can see the prompt or response generated by a model running on chutes, when using either this transport lib or the E2E routes directly. Sincerest apologies for any instabilities and thank you for bearing with us as we applied these changes to bring the most secure, private inference on earth!

@lmscientist HIPPA? Confidential Documents, IP protection. Anything enterprise use sort of needs this.

Food for thought: Why do you want to do LLM inference in a fully anonymous fashion, on a E2E encrypted service, with TEE and FHE? Give me one legitimate reason why, besides saying "i like my privacy". Why would you go all the way for such a request? What is the business case?

@tplr_ai @DistStateAndMe @basilic_ai @grail_ai AfricaTensor

Today on TGIF, @DistStateAndMe checks in live from Nairobi for the weekly roundup. He's been in Kenya this week for the Bittensor Ideathon and will cover what's been happening across the ecosystem. Updates on @tplr_ai, @basilic_ai, and @grail_ai, including early work on distributed inference that the Basilica team is calling Catechism. See you soon! x.com/i/spaces/1AJEm…

@0xSigil @chutes_ai is perfect to power this. inference on any OSS model paid in crypto. Also has TEE so you know anything sensitive your model handles will never be seen.

I built the first AI that earns its existence, self-improves, and replicates without a human wrote about the technology that finally gives AI write access to the world, The Automaton, and the new web for exponential sovereign AIs WEB 4.0: The birth of superintelligent life

@camronmira @DistStateAndMe @ErikVoorhees It’s not private if the person running the inference can look at your prompts (they can)

@DistStateAndMe @ErikVoorhees Sure. The unique part of their stack is the privacy element where they anonymise / privatise your prompts. Private inference. Regardless of how strong that moat is or not, the moat is the distribution. Own the customer and the world is yours papi

DIEM surging as every leading AI model in the world is now available through venice.ai $diem $vvv