Work on yourself until you respect yourself! #bodybuilding #AI #robotics #ReinforcementLearning #solana
somewhere
Joined October 2009
- Tweets1.1K
- Following1.2K
- Followers284
- Likes1.5K
Abi retweeted
Claude Opus 5.5 writes a fugue in the style of Bach. I called Astra’s fugue the most impressive musical feat I’d ever seen from an LLM, and this is just as good. The circle-of-fifths sequence beginning at 1:13 is beautiful. The interplay of voices is natural and satisfying. LLMs are getting to the point where their music is good enough to listen to for pure enjoyment.
Abi retweeted
The most common reasons for failing, in order.
1) Never starting.
2) Quitting too soon.
3) Not doing enough volume.
4) Not doing it consistently enough.
5) Doing too many different things.
6) Not shooting big enough. — @AlexHormozi
Abi retweeted
Most times, people believe that most risks are observed in bold actions. While this may be true to an extent, it’s pertinent to know that, the greatest risk often lies in passive conformity.
That is, the man who follows the crowd may feel safe, but he binds his fate to external forces he cannot control intelligently and wisely at will. And when conditions change as they inevitably do, he’s carried along with them. Unable to adapt because he has never learned to think independently.
Abi retweeted
dark truth about trading no one tells you:
> out of 429k traders on fomo, 24k traders (5%) are in profit
> only 500 traders have made more than $1k ever
> majority have just profited $100 or less
> and 95% of the traders are in a loss
youre trying to compete for that 1% spot guys
stop burning your money
Abi retweeted
Qwen3.8-Flash-Next is here! An open-weight multimodal MoE built on a brand new architecture, with native 256K context extendable to 1M via YaRN. 🤖modelscope.ai/collections/Qw…
🏆 Leads every compared model on SWE-bench Pro (62.5 vs 53.4 for Claude-Opus-4.6 Max), SWE-bench Multilingual, CoWorkBench, JobBench, and Toolathlon Verified.
⚡ 125B params plus 51B n-gram embeddings, only 6B active per token. Stronger than Qwen3.7-Plus on coding and office tasks at roughly 1/9 the training cost. 📚 At 1M context, attention kernels run up to 7.6× faster on prefill and 4.9× on decode.
Abi retweeted
Venice is powering the first robot ever to climb Everest.
Scaling new heights with private, unrestricted inference.
Introducing Pemba
The first robot to attempt Everest
This fall powered by @livekit, with support from @AskVenice , @eastworlds_io and Reflex
Abi retweeted
God was the first vibe coder.
“Let there be light.”
No design doc. No tests. No staging environment. Just shipped the universe directly to production.
Abi retweeted
Harsh truth.
I feel like a lot of people should focus on earning more money.
Investing small amounts per month is great and helps…especially in Bitcoin.
But it will take you absurdly long to get wealthy and retire.
You’re better off betting on yourself to make more money.
- create something
- sell something
- build something
- start a business
Unless you can find a job making $150k+ a year (and get decent raises) inflation will eat you alive.
Of course everyone will point out outliers, but the reality is
You need to find a way to make as much money as possible and invest as much of it as you can.
And betting on yourself can typically get you there faster than climbing a corporate ladder.
Abi retweeted
> be me
> Get a new 4 TB ssd from SanDisk and have Claude backup my home folder
> Opus notices 8 MB/s transfer speed
> Starts investigating, bad cable? What link was negotiated... Ok it's 10 Gbps
> Fun to see all the hypotheses, SLC cache exhausted? SanDisk has that issue....
> Finds two macOS daemons, Spotlight was indexing the backup drive and rsync was reading a virtual mount causing the os to download by cloud drive to "copy" it lol
> Fix both and copy speed is 984 MB/s
Feels good to have a genius looking over my shoulder
Anyone tried Hermes agent for RL workflows? If not, definitely worth a try.
Everything from getting the training data, splitting it, crafting env, initiating the training process, actively watching the training, stopping it, updating rewards, and retaining it. You name it.
Got nvidia/GLM-5.2-NVFP4 running on 8× RTX 6000 Pro after a bit of tinkering. Didn’t find a clean setup anywhere, so had to piece it together.
Right now it’s stable at 450k context with ~70–110 tok/s throughput. Open to suggestions or further optimizations.
#localai #rtx6000pro
Gist with the full setup: gist.github.com/abisheksunil…
Credit where its due: github.com/local-inference-l…
Abi retweeted
Really do not understand in what world will Anthrophic and OpenAI’s market valuation make sense,
When Chinese open source models are consistently three months behind them or by this new benchmark, basically a month.
We’re not the good guys this time, Americans, we are the baddies this time. I don’t understand why anyone would want Dario, Sam Altman, or Donald Trump to be at the steering wheel of AI.
It is objectively great for the world that China is matching America in this race, and that China is choosing the open source path on this.
BREAKING: GLM-5.2 is now 1st on Design Arena.
With an Elo of 1360, GLM-5.2 has jumped ahead of the now unavailable Claude Fable 5.
And it's open weights.
This is an improvement of 4 positions and 27 Elo points to achieve one of the highest Elo scores in our code categories since Design Arena started.
Huge congratulations to the @Zai_org on the release!
Abi retweeted
We are improving constantly the UX of the server to make it easier to use. Now you can run it via a docker image without wrestling with dependencies 🤠
You can now run @luceboxai engine in a container, with one image for every supported GPU.
Until now the only way to run it was to build it yourself: clone the repo with its submodules, install uv, CMake and a CUDA toolkit, wait around 25 minutes while nvcc compiled kernels for six GPU architectures, then download the weights and figure out the right flags.
None of that is hard ofc, but it is a lot of steps if you just want to try the server. There are now prebuilt images on GHCR, one for NVIDIA cards from the RTX 2080 Ti to the RTX 5090 and one for AMD starting with Strix Halo, so the whole thing becomes a pull, a mount and a docker run.
Details below 🧵
Abi retweeted
Today, we’re excited to introduce Miso One, the most emotive voice model in the world.
Miso One is an 8-billion-parameter text-to-speech model for highly expressive speech generation. It emotes like a human and responds faster than a human, with just 110 milliseconds of latency.
We’ve open-sourced the model weights, with API access coming soon.
Hear how Miso One sounds in the thread below.
Abi retweeted
Yes, we did x75 of $gitlawb
bullish on $serv from launch day.. things mooving fast there.
We covered SERV with x40 on $router
I hope you saw next play there..
xitter.cf/thecrypto_dave/status/…
