A.V.
@slckl.bsky.social
Trying to make Rust x AI a reality.
Python survivor, book lover and weird music enjoyer.
Python survivor, book lover and weird music enjoyer.
I love widely used high quality datasets.
November 18, 2025 at 8:14 AM
I love widely used high quality datasets.
Just finished Clevatess, first season. Super solid dark fantasy, with an old school vibe, but enough fresh twists to keep you hooked. Strong personal contender for aoty 2025.
November 11, 2025 at 7:49 PM
Just finished Clevatess, first season. Super solid dark fantasy, with an old school vibe, but enough fresh twists to keep you hooked. Strong personal contender for aoty 2025.
Comparatively tiny models, trained on purpose made data, can into reasoning. Beautiful work!
Breaking: we release a fully synthetic generalist dataset for pretraining, SYNTH and two new SOTA reasoning models exclusively trained on it. Despite having seen only 200 billion tokens, Baguettotron is currently best-in-class in its size range. pleias.fr/blog/blogsyn...
November 10, 2025 at 8:07 PM
Comparatively tiny models, trained on purpose made data, can into reasoning. Beautiful work!
Reposted by A.V.
Surprising: Math requires a lot of memorization
Goodfire is at it again!
They developed a method similar to PCA that measures how much of an LLM’s weights are dedicated to memorization
www.goodfire.ai/research/und...
Goodfire is at it again!
They developed a method similar to PCA that measures how much of an LLM’s weights are dedicated to memorization
www.goodfire.ai/research/und...
November 7, 2025 at 1:02 AM
Surprising: Math requires a lot of memorization
Goodfire is at it again!
They developed a method similar to PCA that measures how much of an LLM’s weights are dedicated to memorization
www.goodfire.ai/research/und...
Goodfire is at it again!
They developed a method similar to PCA that measures how much of an LLM’s weights are dedicated to memorization
www.goodfire.ai/research/und...
Reposted by A.V.
This is the first company that is unironically making actual gynoids and it's not even for porn reasons. They just are.
November 5, 2025 at 4:16 PM
This is the first company that is unironically making actual gynoids and it's not even for porn reasons. They just are.
Reposted by A.V.
Cursor made an LLM
it’s called Composer, it’s an extremely fast model that was previously available under code name Cheetah
it’s an MoE trained in fp8, RL’d on Cursor Agent traces
cursor.com/blog/composer
it’s called Composer, it’s an extremely fast model that was previously available under code name Cheetah
it’s an MoE trained in fp8, RL’d on Cursor Agent traces
cursor.com/blog/composer
Composer: Building a fast frontier model with RL · Cursor
Built to make you extraordinarily productive, Cursor is the best way to code with AI.
cursor.com
October 29, 2025 at 6:39 PM
Cursor made an LLM
it’s called Composer, it’s an extremely fast model that was previously available under code name Cheetah
it’s an MoE trained in fp8, RL’d on Cursor Agent traces
cursor.com/blog/composer
it’s called Composer, it’s an extremely fast model that was previously available under code name Cheetah
it’s an MoE trained in fp8, RL’d on Cursor Agent traces
cursor.com/blog/composer
Reposted by A.V.
Hundreds of hours of European driving data from NVIDIA! 1700 hours total
Big day for autonomous driving research.
Nvidia just dropped 1700 hours of public driving data on HuggingFace from over 2500 cities:
huggingface.co/datasets/nvi...
Nvidia just dropped 1700 hours of public driving data on HuggingFace from over 2500 cities:
huggingface.co/datasets/nvi...
huggingface.co
October 28, 2025 at 8:20 PM
Hundreds of hours of European driving data from NVIDIA! 1700 hours total
New company announced, with the intent of making GPU programming better with Rust: www.vectorware.com/blog/announc...
The founding team has impressive Rust credentials. They're targeting a wide range of usecases, not just ML.
The founding team has impressive Rust credentials. They're targeting a wide range of usecases, not just ML.
Announcing VectorWare
We are building the first GPU-native software company. Today we are sharing the thesis, people, and partners behind it.
www.vectorware.com
October 24, 2025 at 4:54 PM
New company announced, with the intent of making GPU programming better with Rust: www.vectorware.com/blog/announc...
The founding team has impressive Rust credentials. They're targeting a wide range of usecases, not just ML.
The founding team has impressive Rust credentials. They're targeting a wide range of usecases, not just ML.
Reposted by A.V.
codex is definitely "senior engineer" material because it takes forever to think about it before it tells you to fuck off
October 22, 2025 at 6:16 PM
codex is definitely "senior engineer" material because it takes forever to think about it before it tells you to fuck off
Sometimes, when writing throwaway UUID v4s, I feel bad about exhausting the global uuid supply. It can't run out, can it...
October 15, 2025 at 5:03 PM
Sometimes, when writing throwaway UUID v4s, I feel bad about exhausting the global uuid supply. It can't run out, can it...
September 30, 2025 at 10:40 AM
Reposted by A.V.
Reposted by A.V.
Sonnet 4.5
Better than Opus 4.1 on almost every benchmark
Still the classic Sonnet prices, $3/$15
Better than Opus 4.1 on almost every benchmark
Still the classic Sonnet prices, $3/$15
September 29, 2025 at 6:05 PM
Sonnet 4.5
Better than Opus 4.1 on almost every benchmark
Still the classic Sonnet prices, $3/$15
Better than Opus 4.1 on almost every benchmark
Still the classic Sonnet prices, $3/$15
Reposted by A.V.
Alibaba released Qwen3-VL
The flagship model Qwen3-VL-235B-A22B is released as open-weight and available in both Instruct and Thinking versions
✅ Instruct outperforms Gemini 2.5 Pro on key vision benchmarks
✅ Thinking achieves state-of-the-art (SOTA) performance on multimodal reasoning tasks
The flagship model Qwen3-VL-235B-A22B is released as open-weight and available in both Instruct and Thinking versions
✅ Instruct outperforms Gemini 2.5 Pro on key vision benchmarks
✅ Thinking achieves state-of-the-art (SOTA) performance on multimodal reasoning tasks
September 23, 2025 at 11:55 PM
Alibaba released Qwen3-VL
The flagship model Qwen3-VL-235B-A22B is released as open-weight and available in both Instruct and Thinking versions
✅ Instruct outperforms Gemini 2.5 Pro on key vision benchmarks
✅ Thinking achieves state-of-the-art (SOTA) performance on multimodal reasoning tasks
The flagship model Qwen3-VL-235B-A22B is released as open-weight and available in both Instruct and Thinking versions
✅ Instruct outperforms Gemini 2.5 Pro on key vision benchmarks
✅ Thinking achieves state-of-the-art (SOTA) performance on multimodal reasoning tasks
Very cool. More stuff on the bad site: x.com/Alibaba_Qwen...
Alibaba's Qwen3-Omni — the end-to-end omni-modal AI unifying text, image, audio & video in one model
🏆 SOTA on 22/36 audio & AV benchmarks
🌍 119L text / 19L speech in / 10L speech out
⚡ 211ms latency | 🎧 30-min audio understanding
🎨 Fully customizable via system prompts
🏆 SOTA on 22/36 audio & AV benchmarks
🌍 119L text / 19L speech in / 10L speech out
⚡ 211ms latency | 🎧 30-min audio understanding
🎨 Fully customizable via system prompts
September 22, 2025 at 8:11 PM
Very cool. More stuff on the bad site: x.com/Alibaba_Qwen...
Spectral Labs SGS-1: Generate CAD geometry from description, cool. The output is a format compatible with CAD tools and so this seems quite practical.
www.spectrallabs.ai/research/SGS-1
www.spectrallabs.ai/research/SGS-1
Introducing SGS-1
Spectral Labs releases SGS-1: the first generative model for structured CAD.
www.spectrallabs.ai
September 21, 2025 at 8:37 AM
Spectral Labs SGS-1: Generate CAD geometry from description, cool. The output is a format compatible with CAD tools and so this seems quite practical.
www.spectrallabs.ai/research/SGS-1
www.spectrallabs.ai/research/SGS-1
Feels like a JoJo's episode, amazing.
The writers this season aren’t even trying anymore
September 19, 2025 at 6:04 AM
Feels like a JoJo's episode, amazing.
Moondream3 preview dropped, 9B MoE, A2B, visual reasoning, it's beautiful. Claims it's better than big boys opus 4.1, gemini 2.5. pro etc.
@vikhyat.net cooked
moondream.ai/blog/moondre...
x.com/vikhyatk/sta...
@vikhyat.net cooked
moondream.ai/blog/moondre...
x.com/vikhyatk/sta...
September 19, 2025 at 5:57 AM
Moondream3 preview dropped, 9B MoE, A2B, visual reasoning, it's beautiful. Claims it's better than big boys opus 4.1, gemini 2.5. pro etc.
@vikhyat.net cooked
moondream.ai/blog/moondre...
x.com/vikhyatk/sta...
@vikhyat.net cooked
moondream.ai/blog/moondre...
x.com/vikhyatk/sta...
Sometimes you bang against the walls of your own flesh.
www.youtube.com/watch?v=tzvM...
www.youtube.com/watch?v=tzvM...
I Not Me
YouTube video by Osheyack - Topic
www.youtube.com
September 16, 2025 at 7:32 PM
Sometimes you bang against the walls of your own flesh.
www.youtube.com/watch?v=tzvM...
www.youtube.com/watch?v=tzvM...
In addition to some ambient/techno classics pounding the brain, I also find folk music extremely stimulating for doing work. But only on the condition I can't understand the lyrics.
What's this about? No clue
open.spotify.com/track/5B4lRo...
What's this about? No clue
open.spotify.com/track/5B4lRo...
Tuuli
Hedningarna · TRÄ · Song · 1994
open.spotify.com
August 16, 2025 at 9:17 AM
In addition to some ambient/techno classics pounding the brain, I also find folk music extremely stimulating for doing work. But only on the condition I can't understand the lyrics.
What's this about? No clue
open.spotify.com/track/5B4lRo...
What's this about? No clue
open.spotify.com/track/5B4lRo...
Us europoors eating good for once.
Supported Languages:
Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Slovak, Slovenian, Spanish, Swedish, Russian, Ukrainian
Supported Languages:
Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Slovak, Slovenian, Spanish, Swedish, Russian, Ukrainian
Nvidia open-sources (model, data, and code) both speech recognition models and datasets:
- parakeet-tdt-0.6b-v3: blazing fast and accurate ASR inference with PnC and timestamps
huggingface.co/nvidia/parak...
- parakeet-tdt-0.6b-v3: blazing fast and accurate ASR inference with PnC and timestamps
huggingface.co/nvidia/parak...
nvidia/parakeet-tdt-0.6b-v3 · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
huggingface.co
August 16, 2025 at 8:36 AM
Us europoors eating good for once.
Supported Languages:
Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Slovak, Slovenian, Spanish, Swedish, Russian, Ukrainian
Supported Languages:
Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Slovak, Slovenian, Spanish, Swedish, Russian, Ukrainian
Reposted by A.V.
recently gave a talk on <Reality Checks> at two venues, and discussed (and rambled) about how leaderboard chasing is awesome (and we want it to continue) but that this isn't easy because everyone (me! me! me!) wants to write more papers.
the link to the slide deck in the reply.
the link to the slide deck in the reply.
August 12, 2025 at 2:04 AM
recently gave a talk on <Reality Checks> at two venues, and discussed (and rambled) about how leaderboard chasing is awesome (and we want it to continue) but that this isn't easy because everyone (me! me! me!) wants to write more papers.
the link to the slide deck in the reply.
the link to the slide deck in the reply.
Out of all the model drops today, Genie 3 is the most mind boggling. Good job!
Shame all the juicy details are locked down tight...
Shame all the juicy details are locked down tight...
Get ready to enter the simulation...
Genie 3 is a new frontier for world models: its environments remain largely consistent for several minutes, with visual memory extending as far back as 1min. These limitations will only decrease with time.
Welcome to the future.🙌
deepmind.google/discover/blo...
Genie 3 is a new frontier for world models: its environments remain largely consistent for several minutes, with visual memory extending as far back as 1min. These limitations will only decrease with time.
Welcome to the future.🙌
deepmind.google/discover/blo...
August 5, 2025 at 6:20 PM
Out of all the model drops today, Genie 3 is the most mind boggling. Good job!
Shame all the juicy details are locked down tight...
Shame all the juicy details are locked down tight...
Reposted by A.V.
gpt-oss, OpenAI's open weights model
120B & 20B variants, both MoE with 4 experts active
openai.com/index/introd...
120B & 20B variants, both MoE with 4 experts active
openai.com/index/introd...
August 5, 2025 at 5:45 PM
gpt-oss, OpenAI's open weights model
120B & 20B variants, both MoE with 4 experts active
openai.com/index/introd...
120B & 20B variants, both MoE with 4 experts active
openai.com/index/introd...