Omar Garcia (@0xmareveth) header
Omar Garcia (@0xmareveth) profile picture

Founder of Eigna Labs · Building private AI agents · OSS contributor: OpenClaw, NVIDIA NemoClaw, Anthropic buffa, ElevenLabs

EU
Joined May 2026
Photo by Omar Garcia (@0xmareveth)Photo by Omar Garcia (@0xmareveth)Photo by Omar Garcia (@0xmareveth)Photo by Omar Garcia (@0xmareveth)Photo by Omar Garcia (@0xmareveth)Photo by Omar Garcia (@0xmareveth)Photo by Omar Garcia (@0xmareveth)Photo by Omar Garcia (@0xmareveth)Photo by Omar Garcia (@0xmareveth)Photo by Omar Garcia (@0xmareveth)
I keep going back to exploit_agent in GTIG's new report. They found it on an exposed staging server, using frontier and local LLMs. It compares code before and after patches, drafts an exploit and runs it. No human in the loop. First time they've seen that this past quarter.
1
3
The chart goes from a 2025 average of 8 zero-days a month to 11 a month in 2026, then 22 in August. I'd look at that August bar first. The victim telemetry suggests the operator was interested in French media, entertainment and education sites. cloud.google.com/blog/topics…
1
Hugging Face moved Inference Providers onto OpenRouter's router. Same code, same token. The part I like: if a provider fails, the call retries on another one. Before, you got the error. Chart from their post: open-weight went from 26.8% to 71.7% of OpenRouter's top-50 tokens.
1
4
OpenAI signed Apollo Research, METR, Transluce and Goodfire as embedded assessors. I'm more interested in METR's piece: OpenAI's largest RL run to date, looking for misaligned behavior that existing tests may have missed. Access depends on the job. I'd wait for the reports.
1
24
Confidential access is on site, on OpenAI devices. They publish on their own. OpenAI can redact confidential information and ask for disclosure delays. Apollo and Goodfire are paid. METR and Transluce aren't, but may take API credits or compute. openai.com/index/our-first-p…
26
Qwen4-Max: 43 on the Artificial Analysis Intelligence Index, $3.27 per task. Claude Fable 5.2 (Max): 65, $26.54. The line I'd read before swapping it into my agents is the token count. It used 127M tokens on the index, median is 78.6M. So I'll price it on my own traces first.
1
27
Independent numbers are on Artificial Analysis. Qwen's own claim, from internal evals, is up to ~50% fewer tokens than Qwen3.8-Max. They also promise open weights for smaller Qwen4 models, which are the ones I'd run locally. artificialanalysis.ai/models…
20
Ethan Mollick has a post up on SatBench. First chart: 63% of the 173 AI benchmarks labs cite most are solved, 109 of 173. The cutoff is 85% on the best AI score. I'd start with the footnote. The 63% includes labs' own launch posts. Counting only independent runs, 24% are solved.
1
1
14
Newer benchmarks also fall faster. The share solved within a year of release is 17% for 2023 tests, 30% for 2024, and 43% for 2025. Humanity's Last Exam is the top holdout at 74%. Data from Benchmark Almanac, as of Nov 14. teapawt.org/emollick/status/212175…
AI is getting better faster than our measurements are I threw together SatBench, which measures how quickly major AI benchmarks become saturated (effectively beaten). Of 173 benchmarks cited by the Labs since 2023, 63% are now saturated. Saturation is happening faster each year.
10
llama.cpp and ggml both reached v1.0.0, three and a half years after llama.cpp started. They will bump the major version when the public C API breaks. New features and model support land in a minor release, and a bug fix lands in a patch.
1
10
The versioning details are in the llama.cpp team's post on Hugging Face. They also put coverage at more than 150 architecture variants, plus an XCFramework for iOS and macOS. huggingface.co/blog/ggml-org…
5
People Make Things has a post up on SUS-Bench. 250,000+ unpaid people from 167 countries played 500,000+ rounds of a mafia-like game with and against AI. The thread puts 7 of 28 models above the human baseline, 4 between people and a simulated fly brain, and 17 below the fly.
1
8
The scoring details are on the SUS-Bench page. The part I keep going back to: they compare how often a model's team wins with the benchmark bots. The launch thread says a simulated fly brain controls one player and its neurons decide when to walk and turn. peoplemakethings.com/sus
3
OpenAI has a post up on ChatGPT Wallet. A dot can complete a purchase after you review the items, total and selected card, then approve with Face ID or a fingerprint scan. The token comes from the payment network. The business and the agent don't get the actual card details.
1
16
OpenAI also published a Commerce Partner Playbook. It tells sites to recognize signed requests using Web Bot Auth, and says risk systems may still flag legitimate purchases made by agents. openai.com/index/introducing…
9
Edge0 released weights for a 145B MoE. Their laptop tests report 29 GB peak memory and over 80 tokens/s decode, with about 10B active per token. Hot experts stay in the GPU. An expert library of about 40 GB sits in system RAM. Vocabulary data is read off the SSD.
1
1
33
Read the licences first. The model card lists Qwen3.8-Flash-Next as the base. Speed numbers are Edge0's own tests. Weights use the Edge0 Community Model License plus Qwen's; a commercial model service needs a separate license. Framework code is Apache 2.0. huggingface.co/Edge0/Edge0-1…
33
Reading NVIDIA's GR00T 2 page. The part I keep going back to: a reasoning agent plans the steps and calls GR00T 2 as a tool through an MCP server, and GR00T 2 does the manipulation. Toy assemblies: 98, 97 and 97 of 100 with the agent, 71 to 84 without.
1
1
13
The other number I liked: on 50 tasks the target robot had no data for, 40% success zero-shot. Five demos per task and a 30 minute LoRA fine-tune on one RTX 5090 got it to 95%. research.nvidia.com/labs/gea…
8