Brian Peña Calero (@brianmsm) profile picture

Studying Máster Metodología en Ciencias del Comportamiento y de la Salud - UCM. Interest in mental health issues, psychometric and empirical research.

Madrid, Spain
Joined January 2014
Photo by Brian Peña Calero (@brianmsm)Photo by Brian Peña Calero (@brianmsm)Photo by Brian Peña Calero (@brianmsm)Photo by Brian Peña Calero (@brianmsm)Photo by Brian Peña Calero (@brianmsm)Photo by Brian Peña Calero (@brianmsm)Photo by Brian Peña Calero (@brianmsm)Photo by Brian Peña Calero (@brianmsm)Photo by Brian Peña Calero (@brianmsm)Photo by Brian Peña Calero (@brianmsm)
Brian Peña Calero retweeted
After @BellularGaming's take on a native @t3dotcodes addon for World of Warcraft instead of my overlay approach, I had to take a swing. This is better in every way. It feels totally native, and with full controller and dictation support I can drive threads from the couch.
39
31
29
655
173,903
Brian Peña Calero retweeted
Save the date! November 24th State of Open Models is coming 👀👏 Extensive research about open models ecosystem. What you can expect? An analysis of the Open Models actual moment based on data from the most important companies right now in the field - What models are the most downloaded. And the most used - How closed are open models to closed - How open models are used - Economics of open models - Who are using them - Where are running Just 4 days for the release!
19
6
3
39
3,324
Brian Peña Calero retweeted
Excited to launch Gemini 4 Neon. It's cheaper, faster, and better behaved than Carbon. It beats Opus 5.5 and Sol 6.2 on most benchmarks and #3 on AA. It's been great to use internally so excited for you to try it on Antigravity today on all paid plans! Please try before GTA 6 🙂
280
147
70
3,459
323,027
Brian Peña Calero retweeted
Una sola persona acaba de jubilar a los bootcamps de programación. Hizo una web que enseña con animaciones cómo funcionan por dentro las apps con millones de usuarios. Un ejemplo: una estrella del pop con 50 millones de seguidores publica algo. Su publicación tarda más de 8 minutos en llegar a todos y, mientras tanto, las publicaciones de los demás se quedan esperando. La lección te enseña cómo bajar esa espera a unos 2 segundos. 40 lecciones así, animadas. Gratis, sin cuenta y en español. Lo que antes era un bootcamp de miles de dólares ahora parece una pestaña del navegador.
12
254
1
1,768
75,631
Brian Peña Calero retweeted
Qwen4-Max, the new flagship model from Alibaba, is free in Nous Portal for 5 days in partnership with @inferact and @Alibaba_Qwen. It’s built for long-horizon work, with native vision and 1M context. Try it today at portal.nousresearch.com
181
202
113
2,663
651,680
Brian Peña Calero retweeted
GPT-6.2 Astra is the new #1 on the LLM Creative Story-Writing Benchmark! 🏆 It scores 4.30, ahead of Claude Fable 5.2 at 3.56. Against its predecessor, GPT-6 Astra, it swept all 50 matched prompts. Qwen 4 Max jumps 3.72 points over its predecessor, from -0.17 for Qwen 3.8 Max to 3.55. That puts it at #3 overall, just behind Fable 5.2. 🥉 ByteDance Seed2.1 Pro scores -0.67, up from -2.18 for Seed2.0 Pro. GPT-6.2 Astra is better at letting the reader discover what a story means. GPT-6 Astra often builds toward a familiar turn: the protagonist recognizes a mistake, acts to put it right, and names the lesson. In 6.2, several characters can shape the outcome, and what one of them does late in the story can change how we understand an earlier scene. Its endings need less explanation, and there’s more room for mixed motives, surprise, and humor. Astra 6.2 has recognizable habits of its own. It gravitates toward family ties, skilled handwork, and the things people learn by watching someone they love. It often ends with a small, practical gesture and trusts the reader to see what has changed between the characters. That restraint gives its best endings emotional weight. Across many stories, though, the pattern becomes familiar. The plots can still surprise, even as very different briefs lead back to the same intimate, old-fashioned world. Each brief asks for a 600-800-word story in which ten required elements shape what happens. Three judges from different model families evaluate each new pair in both presentation orders, excluding the writers’ families. The leaderboard now covers 67 writer models, 134,964 judgments and 20,589 matched story pairs github.com/lechmazur/writing…
18
18
2
209
9,981
Brian Peña Calero retweeted
FrontierHarness v1.1 results: when you switch models, consider switching harnesses too, at least if you care about your token bill. The full run: 83B tokens across 27,280 agent runs, each in its own @runta runtime, inference by @baseten. Codex CLI ranks 1st on cost with GLM-5.3, and 10th of 22 with Kimi-K3. On any one model, cost per pass varies by up to 4.4x across harnesses. Only Pi never left the top 6 on cost. 🧵
42
20
14
162
25,852
Brian Peña Calero retweeted
Introducing Qwen3.8-27B-pi-adapted 🤗 We adapted Qwen's model specifically for the pi coding agent using reinforcement learning and a LoRA adapter. The results, measured as pass@1: 🚀 SWE-bench Pro (Hard 51): 29.4% → 62.7% 💻 Terminal-Bench 2.1: 39.5% → 55.8% ⚡ 55–62% fewer tool calls per task This is the power of open models: you can adapt them to YOUR tools and YOUR use cases. Let’s build! 💛 Open weights: huggingface.co/harness-hub/Q… Methodology: huggingface.co/blog/harness-…
310
647
175
5,999
458,028
Brian Peña Calero retweeted
And here we go…GPT-6.2 Astra and GPT-6.2 Sol are yours to make more games with. Oh, and they’re also super efficient and really good at making slides, building websites, and tackling the work you were doing before you started making games. Rolling out today to ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu users. openai.com/index/introducing…
🤖 Made with AI
591
1,184
450
13,138
977,878
Brian Peña Calero retweeted
If you don't wanna deal with abysmal/confusing graphs OpenAI put on, here you go for the number they have posted in that tweet. Clear Anthropic style!
Some of you have been asking very politely. Some of you have just been asking. GPT-6.2 Astra and Sol are now available. Super efficient and built to perform where it counts—the real world. openai.com/index/introducing…
12
6
2
56
44,730
Brian Peña Calero retweeted
NVIDIA ha lanzando su propia herramienta para comprobar la calidad de las skills de agentes y medir sus resultados. → github.com/NVIDIA/SkillEvalu…
9
115
2
870
27,542
La flexibilidad que tiene Simyo para configurar la tarifa que quieras #simyoGooglePixelMola
¡SORTEAMOS UN GOOGLE PIXEL 11 PRO! ¿Lo quieres? 😉 - Sigue a @simyo_es - RT a este Tweet - Dinos lo que más te gusta de simyo y de este móvil usando #simyoGooglePixelMola Hasta el 15/11 a las 23:59h. Bases: ow.ly/lwR35104vJs
22
Brian Peña Calero retweeted
We’re spending up to 100B tokens across 26,040 agent runs to answer one question: when you switch models, should you switch harnesses too? Our v1.0 FrontierHarness leaderboard ranked harnesses on a single model. V1.1 expands it to 21 harnesses × 5 models × 62 tasks × 4 attempts, 72x the size of v1.0. Results are coming in live. Link below🧵
91
65
75
666
498,757
Brian Peña Calero retweeted
¿Recuerdan REA, el proyecto que hacía ingeniería inversa a cualquier cosa? Un agente de IA al que le dabas un binario, una app de Electron o un programa compilado, y te devolvía el código. En la práctica significaba que cualquier programa que exista podía ser destripado y reconstruido. Bueno, acaba de llegar su antídoto. Se llama BLA, Blind LLM Agents, y pasó de 0 a casi 4,000 estrellas en GitHub en mes y medio. La idea es simple: en vez de ocultar tu código, envenenas el análisis. Siembra claves falsas, lógica señuelo y trampas de recursos dentro de tu build. El usuario normal no nota nada; el agente de IA que intenta destripar tu app quema tokens siguiendo pistas falsas, hasta que termina “entendiendo” algo que no existe. Según el proyecto, el atacante gasta entre 300 y 2,800 veces más tokens. Son cifras propias, nadie las ha verificado todavía. Pero la señal es clara, la ofuscación clásica ya no alcanza; ahora proteges tu código pensando en cómo lee una IA Es open source, MIT, y está en GitHub github.com/qybaihe/bla
40
239
10
2,145
88,330
Brian Peña Calero retweeted
1/ announcing our smallest, most efficient little model is here: muse ember 🔥 built for work that agents repeat all day, it gets you more intelligence per dollar than Haiku, Luna, and all other flash models at any reasoning setting and does it at a third of the cost or less!
290
340
150
6,139
401,627
Brian Peña Calero retweeted
github.com/mmnavarr/overseer is the must have plugin for Tern
5
11
1
190
16,479
Brian Peña Calero retweeted
My old Echo Dot run Hermes now, instead of Alexa. I built a Home Assistant integration that connects it to a Hermes profile, with its memory and tools. Code, setup skill and the hardware recovery notes are public. github.com/majorminorlabs/he… @NousResearch
Please definitely dm me or send me your feedback - I reaaaaaaaaaaaaaaaaaaally wanna replace Google Home/Alexa for people one day.
99
102
31
1,251
192,594
Brian Peña Calero retweeted
opus'a ucuz bir işçi eklemek işi ucuzlatmıyor. reddit'te bir kişi 12 zor görevi ölçtü. opus tek başına %81 başarı oranı ile görev başına 1,20 dolar harcadı. opus planlayıp deepseek v4.1 flash kod yazınca %67 başarı oranı ile 1,29 dolar harcadı. 7 takım tasarımı, 41 kurulum, yaklaşık 1.200 koşu ile test edilmiş. opus'un başında olduğu en iyi takım %69'da kalmış. sebep model seçimi değil görevler arası devir teslim. worker kod yazarken opus'un prompt cache'i düşüyor ve opus aynı contexti yeniden okuyor ve tekrar maliyet çıkartıyor. bu kalem tek başına worker'ın maliyetinden fazla çıkmış. aynı tabloda gpt-6.1 terra %78 ve sol %75 başarı oranı ve görev başına 0,17 dolar olarak ölçülmüş. genel olarak 12 görev kesin bir karar vermek zor ve küçük farklar gürültü yaratabilir. ama 41 takımın hiçbiri tek opus'u geçmemiş.
24
17
6
194
28,825
Brian Peña Calero retweeted
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math. rikiyat.github.io/blog/weight-sp… github.com/RikiyaT/lm_und… huggingface.co/Qwen/Qwen3.5-0… huggingface.co/RikiyaT/Qwen3.…
2,367
7,254
5,755
55,906
10,097,144
Brian Peña Calero retweeted
Did you know the same intelligence can cost 10× more, need 6× more output tokens and take twice as long? In addition to task spend, Agent Arena now shows which agents give you best performance for every output token generated and every minute you wait. These real-world usage plots reveal that costs grow most steeply with intelligence, waiting longer pays off unevenly, and output tokens barely predict the bill. To find the best model for you, see the link in thread.
30
7
5
226
29,434