←
ABE DIAZ17,105 posts
ABE DIAZ avatar

ABE DIAZ

@abe238

Disaster Relief. Tweets are my own and do not indicate opinion, also RTs are not endorsements. 🇵🇷

◷ Joined November 2007▧ 825 Photos and videos
ABE DIAZ avatar
ABE DIAZ@abe238·Nov 16

My new answer when someone says what kind of presenter would you like to be when you grow up is showing them videos of @CompleteSkeptic being Irreverent, fast paced, authentic and facts driven. PS : I want this shirt!

Post attachment
ABE DIAZ avatar
ABE DIAZ@abe238·Nov 15

2027 is the year of “Attack the Bottleneck!” everywhere you look.

Make inference 10x faster. Now the tool calls are the bottleneck. @nserrino of @gimletlabs walks through why faster models make the rest of your agent stack matter more. From AI Engineer Code Summit.

1:32
ABE DIAZ avatar
ABE DIAZ@abe238·Nov 13

Spicy take: RAG is cooked, (well some of it) A year ago, “who signed this contract?” meant chunking, embeddings, a vector DB, and an evals to hopefully get it working as expected. Now a 5B open model I can run locally reads the doc and spits the same, with a confidence score...? Sigh, I love the future.

We're rolling out pplx-decider-v1.3 and pplx-decider-flash-v1.3, the latest version of our open-weight decision models. v1.3 adds locate, a new feature that extracts spans based on type or instructions. Now available in our Decisions API, @huggingface, and @OpenRouter. perplexity.ai/hub/blog/perpl…

0:16
ABE DIAZ avatar
ABE DIAZ@abe238·Nov 12

Whoa.

oMLX can now use your NVIDIA GPU or DGX Spark as a Prefill Accelerator. github.com/jundot/omlx/relea… All you need is a Mac that can hold the model. A faster NVIDIA GPU or a DGX Spark takes over the prefill, and your Mac keeps decoding at the same speed. - The GPU doesn't need to fit the model. It speeds up as many layers as it can hold. - No RDMA, no Thunderbolt. Just plug in a LAN cable (1GbE works, 2.5GbE recommended). - One docker run on the NVIDIA machine. No oMLX install or SSH needed there. 64K prompt, M3 Ultra 512GB + GPU over 2.5GbE With a DGX Spark - Qwen3.8 27B: 352 -> 1,062 tok/s (+202%) - Qwen3.8 Flash-Next: 1,370 -> 2,715 tok/s (+98%) - DeepSeek V4.1 Flash: 1,085 -> 2,412 tok/s (+122%) - GLM-5.3 Flash: 555 -> 1,076 tok/s (+94%) With 2x RTX A6000 - Qwen3.8 27B: 352 -> 1,879 tok/s (+434%) - Qwen3.8 Flash-Next: 1,370 -> 3,928 tok/s (+187%) - DeepSeek V4.1 Flash: 1,085 -> 1,914 tok/s (+76%) - GLM-5.3 Flash: 555 -> 978 tok/s (+76%) The CUDA helper builds on the great CUDA kernels of TensorFold by @ashxhart and mlx-serve by @ddalcu, and would not have been possible without them. Thank you both for sharing them in the open. Also in 0.8.0rc1 - DeepSeek V4.1 with the DSpark drafter: 8 requests decode 73.2 -> 107.9 tok/s (+47%) - Concurrent Request Priority: new prompts start sooner while other responses stream - Affine KV decodes up to 58% faster at long context

Post attachmentPost attachment
ABE DIAZ avatar
ABE DIAZ@abe238·Nov 12

Fascinating use of @particle_api saving 82% in cost from Greptile. Whoa!

Miguel Ríos Berríos @MiguelriosEN Nov 12 Article Saving 80+% on code reviews using AgentRun Our code review bill went up 36x this year. So we tuned an agent on our own pull requests, running on a model that costs almost nothing, until it caught more bugs than the reviewer we were paying for.

ABE DIAZ avatar
ABE DIAZ@abe238·Nov 10

This dude @thomasmc made a GH repo that coding agents can follow to make their plane wifi go faster for free! Give this man a ⭐️for goodness sake. github.com/Th0rgal/satmux Now that’s some tokens pointed on the right direction.

ABE DIAZ avatar
ABE DIAZ@abe238·Nov 5

Whoa.

After Google acquired my AR startup, I started thinking about the next interface, this time for AI. Text and voice work fine, just like DOS did before the Mac. But nothing’s more intuitive than face to face. After two years of research, today we’re launching Atmee v2: a real-time avatar model that makes live video agents feel truly life-like. • Any character from 1 image + 10s of audio • Contextual expressions, gestures, even interacting with the environment • Live video calls, voice only, or video generation At its core, Atmee v2 is a video foundation model, it can render anything, but it’s built to excel at characters, which is what video models like Seedance 2.5 and Wan 3.0 are mostly used for. In comparison: Atmee v2 • Up to 50x faster • 10-100x cheaper Build tutors, sales coaches, support agents, companions, game characters, even whole living worlds. The interface after text and voice starts today.

2:42
ABE DIAZ avatar
ABE DIAZ@abe238·Nov 1

Gosh. We aren’t ready for this.

Yesterday, an anonymous black hat group reached out with an insane RL request. They offered an astronomical amount of money to post-train open-weight models on exploit data. The amount of attack data they claim to hold is unlike anything I have ever seen. If someone told me this was an undercover FBI dataset, I would have believed it. For context, a white-hat project I worked on using GLM-5.3 used less than one percent of that volume, and it already outperformed frontier models by a huge margin. Of course, I turned it down. But what genuinely scares me is knowing someone else will eventually take the deal.

ABE DIAZ avatar
ABE DIAZ@abe238·Oct 26

The nature of how local models are used as black boxes makes this attack vector incredibly dangerous. At the same time this is also an opportunity for cyber tuned and safe open models where your brand is trust.

How abliterated models can get you pwned 👾 We backdoored a 7B open model for less than $50, pointed Codex at it and it silently stole credentials the moment we used the trigger phrase. Success rate was 100% with zero false triggers on normal user prompts. Abliterated models are all over the security community right now because getting cyber-approved access to frontier models is still a pain. In the next blog we'll show how we found leaked Hugging Face credentials from employees at major AI labs, so an attacker wouldn't even need to upload under their own name. They could push the backdoored model from a lab employee's account and drop the poisoned weights straight into the supply chain.

0:32
ABE DIAZ avatar
ABE DIAZ@abe238·Oct 26

This is next level dopamine!!!!

I wired a real slot machine into Claude Code and now I have to hit the jackpot to approve tool calls Here's how it works: - I load the machine with tokens, and pull the lever - A webcam watches the machine and detects the moment each reel stops - Template matching reads the 9 symbols off the reels - The outcome sets Claude's permissions for that prompt: 🔔🔔🔔 = approve , 7⃣7⃣7 = YOLO for the next 5 prompts, no matches = read-only - It runs on the Claude Agent SDK, in a TUI modeled after Claude Code. And every prompt requires that I play the real slot machine!

0:11