Tiger

If I were an actor, @HP would have replaced this in milliseconds. Because I am just a student with an electrical injury, I get 81 days of harassment. Is this the 'Equity' you promise @narendramodi @AshwiniVaishnaw? Case 5156493808 #Discrimination #DigitalIndia

@henryzhongsc @brianbellx @tianyi___zhang I have one doubt that as we both know huff llm or dfloat reduce size losslessly by reducing weights by compressing them down to 70 percent value, there's one thing evil here going on speed reduces 2x which is bad though. Are there any methods,papers or formulas with speed

This looks like the README you updated after my initial reply, and — as per my last comment — I find it somewhat shallow. I’ll elaborate a bit here, and please do excuse me for sounding pedantic. A cleaner lineage, to me, is: Liguori arxiv.org/abs/2404.10896 (which neither you nor DF11/ZipNN cited) then ZipNN identified the exponent redundancy and developed it into a storage compression trick. NeuZip and DFloat11 then supported GPU inference, with DF11 being much more general-purpose. I’d wholeheartedly recommend the Extended Related Works section (Appendix B) of DF11 if you haven’t read it already. NF12 (which you did not cite) and ZipServ then replaced DFloat11’s variable-length encoding with fixed-length encoding, with ZipServ already using a fused kernel. There maybe more, but we should be able to track the all by iterative through all works that cited DF11. Your writeup does not clearly explain these prior works, and frankly, I am not seeing the unique contribution if exponent compression, fixed-length encoding, and fused inference have already been explored together. Arguing that it supports newer models is not too meaningful by itself, as the redundancy has already been established to be a BF16 characteristic. I also do not think I did a perfect job documenting this lineage in DFloat11 for missing Liguori. With the traction your post is getting, this is a good opportunity to correct the record. Properly discussing the prior art and providing comprehensive evaluations are things that we should all be striving for.

I removed 423 GB from GLM‑5.2 without changing the model. 1,403 GB → 980 GB. 753B weights. Bit for bit exact. No quantization or retraining. The weights remain compressed in VRAM instead of rebuilding the full model first. Full writeup and repo in the next post.

@brianbellx any speed reduction on like we see in i dfloat 11 ? @brianbellx

@HarryTandy @DAIEvolutionHub Not for me. I am not even subscriber. All the best bro.

@DAIEvolutionHub auto mode is an absolute game changer for laziness

Claude Code team just dropped a free course on loop engineering with Fable 5: 00:00 - how Claude Code works under the hood 05:01 - the agentic loop explained 16:21 - the feature 99% of devs miss: auto mode 19:01 - why voice beats typing 32:34 - auto code review with draft PRs 58:39 - Fable 5 for non-code work this free course replaces every paid Claude Code tutorial watch today, then read the article below on loop engineering by Karpathy

@viditchess @Money_sh_ Had you analysed 90 percent of most common moves of your opponent?

@Money_sh_ I am doing all this to forget that pain! :)

Didn’t see this coming! Pleasant surprise :)

@viditchess @Money_sh_ I don't think I or anyone understands your pain .No one has shouldered the responsibilities which you faced because we have different environments. We only see you as a professional GM who is earning high while forgetting reality is our mistake.

@viditchess @Money_sh_ Yeah I think it's way more tough for you to keep your position when you have responsibilities of a Grand Master and you have folks to challenge your existing dominence. Though not your fan but I use different thinking patterns. All the best .

@emmanuel_belo @openclaw @TencentHunyuan @OpenRouter @ollama @Zai_org Meanwhile silly claude tells users " Use caution before running this prompt. Malicious conversation content could trick Claude into attempting harmful actions or sharing your data."

@tiger43026 @openclaw @TencentHunyuan @OpenRouter @ollama @Zai_org Nah, they're not, but the subscription already comes with a decent tier.

Tencent Hy3 from @TencentHunyuan is free on @OpenRouter through July 21. 295B MoE, 256K context, built for coding, reasoning, agents and reliable tool use. Try it in OpenClaw today: openclaw models set openrouter/tencent/hy3:free

@emmanuel_belo @openclaw @TencentHunyuan @OpenRouter @ollama @Zai_org With cheaper price like beating opus 4.6 ,4.7 . 4.8 and fable are left but how long is something that makes these things actually interesting. They went from something to whatcwe see today.

@emmanuel_belo @openclaw @TencentHunyuan @OpenRouter @ollama @Zai_org Yeah glm is underated too though. If glm 5.3 is out then who is gonna care for fable 5 or any new anthropic ai because glm is giving too much super powers

@emmanuel_belo @openclaw @TencentHunyuan @OpenRouter @ollama @Zai_org Are tokens unlimited

@openclaw @TencentHunyuan @OpenRouter I'm already doing great with @ollama and @Zai_org GLM5.2 on my $20/month subscription, so I don't really need a limited free tier.

@eire1274 Or micro one

@eire1274 What's that mini pc . Please refer me link . I think you are practical person

the hardware pays for itself faster than the cloud bill it replaces. a full 4-node cluster, 128gb ram, 32 cores, 4tb ssd, under 30W.

@askkaruly Were you using dev tools

What I've done😭😭 I just hacked the system to get unlimited tokens. Basically the token usage was calculated locally and then send to the server so I just manipulated with that. please comment and share this post so I can at least say that this was for conent...

Incubator CEO announced challenge to use maximum amount of tokens and top 10 will get Claude Max. Please help me out and drop your comment to maximise the token usage...

@heyshrutimishra Not a great tool .Hype is more than its usefulness. Dumb project.

PewDiePie stopped making videos and built a free, open-source alternative to ChatGPT. It is called Odysseus. 80,000 GitHub stars in few days. It’s complete workspace that runs on your machine: AI chat, autonomous agent with file access and memory, email client, notes, research tools, calendar. No cloud. No subscription. No data leaving your machine. He built it with 8 RTX 4090s because he did not want his data on someone else's servers. Android Authority called it surprisingly brilliant.

@elliotarledge How much was accuracy loss ??

Claude Fable 5 [max] wrote the first genuine (and fastest) megakernel ever submitted to KernelBench-Mega. It was tested on: Kimi-Linear W4A16 batch-1 decode for RTX PRO 6000 Blackwell. Every prior model "won" it with a multi-kernel Triton pipeline that fails our single-fused-kernel authenticity gate > Opus 4.8 at 14.4x > GLM-5.2 11.1x > GPT-5.5 4.3x > Sonnet 5 4.0x. Fable shipped 18.7x over reference, and torch.profiler shows exactly ONE cooperative kernel launch per decoded token. Int4 dequant (nibbles unpacked in-register, never materialized), conv+SiLU, KDA gated-delta state, MLA absorbed-latent attention with online softmax, MoE router + top-8 experts, RMSNorms, even the KV cache append all inside one launch, staged by 14 grid barriers. We overwrote its input buffers mid-audit to prove it recomputes on live data. It does. The advantage grows with context. 17.8x at 2k, 18.9x at 8k, 19.5x at 16k. Longer context means a bigger KV cache and more attention work per token which is usually where a decode kernel bleeds. Keeping everything in one launch amortizes the fixed barrier overhead and the int4 GEMV stays bandwidth-bound, so the gap over the reference widens instead of closing. It spent 64% of the session in silence timing the baseline, microbenchmarking grid barriers, deriving a ~29x bytes/token roofline, then wrote the whole kernel once, hit 14.4x on the first benchmark, and spent the last hour deleting barriers and making int4 dequant free (one LOP3 + HSUB2/HMUL2). The one regression it tried (finer split-K) it measured and reverted instead of rationalizing. kernelbench.com/mega

@elliotarledge How much is sponsership cost in general you charge? I am curious to know. I think that since you are such renowned dev,you even would be willing to get massive sponsership payment.

@askkaruly Like I need some extra benifits.I literally belong 0 merit background and i am self taught devloper so I face many constraints.

@askkaruly Do you have any reach for cloud gpu providers ? Need thy help if you can arrange?

I was rejected and I am happy for that because at this point I am tired seeing "App Under Review". It took 7days btw

@askkaruly Nah !! It's kinda killing coders . Claude Fable is handicapped in front of those programmers who are literally leveraging open source AI models to build their best models to rival claude while having limited compute. Let's see how long can they live .

Fable 5 is back alive. Get claude max and build whatever you want. It literally one shots everything!!! btw i think programmers are cooked

@PeytonAGI @ClementDelangue @PalantirTech Am I right??

Lots of people are advocating for more American open-source models these days which is amazing but very few people do anything about it! Latest example, Alex Karp came out advocating for American open-source models as a necessity! At the same time, @PalantirTech is a free org on HF with 0 open-source models and 0 public datasets shared. Time to switch from talking to contributing for all!