@alextoussss why can we not use those shooting lasers to get them killed? Isn't a drone overengineering here?
English
Sven Goerdes
131 posts
Extremely excited to announce our first air-to-air kill of a flying moth by an autonomous micro-drone. This is a big step towards completely eradicating mosquitoes.
English
YOU’RE MISERABLE BECAUSE YOU’RE NOT FUNMAXXING
HAVE YOU EVER CONSIDERED BEING YOUR TRUE SELF
HAVE YOU EVER THOUGHT ABOUT INDULGING IN WHAT MAKES YOU HAPPY
HAVE YOU TRIED BEING KIND TO SEE THAT ITS POSSIBLE TO SMILE
THIS REFLECTS ON THE OUTPUTS OF YOUR AGENTS BTW
English
@cleoabram @TIME Well deserved Cleo! Please keep your enthusiastic and curiosity driven videos coming! It is hard to find such a positive attitude!
English
Excited to be part of TIME100 Creators! HUGE* is just getting started, and we're in incredible company on this list. Thank you @TIME! ti.me/100creators
English
Holy crank, need to do some more testing.
@PrismML, since this is a clear proof of concept, could you, in theory, do the same with something massive like a GLM 5.2 base and get it to run well on attainable consumer hardware? Incredible work here, I am looking forward to trying it!
PrismML@PrismML
Today, we’re announcing Bonsai 27B: the first 27B-class model to run on a phone. Bonsai 27B is the new multimodal flagship of the Bonsai family. Based on Qwen3.6 27B, it brings a new capability tier to local AI: multi-step reasoning, structured tool use, long-context workflows, and coherent agentic loops. Until now, models in this class have been impractical to deploy locally. A 27B model occupies roughly 54 GB in 16-bit precision, and even a strong 4-bit build is around 18GB - too large for a phone and for most laptops. Bonsai 27B changes that. It comes in two variants: • Ternary Bonsai 27B: 5.9 GB, 1.71 effective bits per weight, optimized for laptop-class quality. • 1-bit Bonsai 27B: 3.9 GB, 1.125 effective bits per weight, optimized for phone-class footprint. Everything is open-sourced today under the Apache 2.0 license.
English
Sven Goerdes retweetledi
Fully support this important proposal from @demishassabis. The time for us all to act is now.
"...we must use this precious window before AGI arrives to shape this technology for the benefit of all humanity."
Demis Hassabis@demishassabis
English
@8teAPi While I agree that AI capability is doubling extremely fast, you still have to keep in mind that they are tracking software tasks and not general tasks. Also as others mentioned, they are in uncertainty territory regarding their measurements.
English
@KingBootoshi ok haha, I'm curious what your neurodivergency will brink to life. Lets goo
English
@GoerdesSven it’s the universe i have in my mind and i bring it to life with ai
English
finetuning the gemma 12b base model to represent the Boo universe entirely
the goal is to truly create a self-evolving story that lives and grows in real time
every checkpoint is an archive of its past
from here i then use this universe-model as the base model for finetuning special characters that know the FULL lore
by finetuning the WORLD MODEL itself into the specialized character
i get a neural net that completely knows the lore and story of the world + has its own personality and perspective on it!
i'm truly creating characters that are a living embodiment of both personality and lores, completely baked into neural nets, running locally on my MACBOOK
this was actually one of my dreams and I nearly side-quested it to full completion in a week purely through Fable LMFAOOOO
English
Sven Goerdes retweetledi
Ich habe heute versucht, die Chatkontrolle zu stoppen.
Am Wochenende mussten Sibylle Berg und ich Parlamentspräsidentin Metsola schriftlich mitteilen, dass das Durchprügeln der Chatkontrolle im Eilverfahren leider gegen die Geschäftsordnung des EU-Parlaments verstößt. Während Metsola der interessierten Presse daraufhin (fälschlicherweise) mitteilte, dass das schon alles seine Richtigkeit habe, warten wir immer noch auf ihre Antwort.
Deshalb wollte ich es ihr heute bei der Eröffnungssitzung in Straßburg noch einmal erklären. Und musste erstaunt feststellen, dass die Präsidentin doch Regeln kennt: Exakt nach 60 Sekunden hat sie mir das Mikrophon abgestellt. (regelkonform, wird aber selten gemacht)
Dabei hätte ich noch einiges zu sagen gehabt: "Frau Präsidentin, Sie wachen nach Artikel 22 über dieses Regelwerk (mit der ausgedruckten Geschäftsordnung wedelnd) - erklären Sie den Eilantrag für unzulässig. Wir sind hier schließlich nicht auf Malta! Die aktuelle Fassung der Geschäftsordnung überreiche ich Ihnen gerne persönlich. In der MEP-Bar."
Wie es nun weitergeht? Morgen wird über das Eilverfahren abgestimmt, obwohl diese Abstimmung gar nicht stattfinden dürfte. Wenn der Antrag erfolgreich ist, kommt die Chatkontrolle Donnerstagmittag zur Abstimmung ins Plenum. Um sie noch zu stoppen, müssten 361 Abgeordnete - eine qualifizierte Mehrheit - DAGEGEN stimmen.
Die schlechte Nachricht: Donnerstag ist der letzte Tag vor der Sommerpause und viele MEPs dürften bereits auf dem Weg in den Urlaub sein. Ein Schelm, wer Böses denkt bei dieser Terminierung... Smiley!
Wenn die Chatkontrolle durchgeht, dürfen die Plattformen (also die US-Tech-Bros) wieder & weiterhin fröhlich & ganz legal Ihre Nachrichten scannen. Schreiben Sie also gefälligst etwas unterhaltsamer in den kommenden Wochen... ZwinkerSmiley!
Deutsch
@deckard_the_dev @Polymarket Yes, but when you look at for example Gemma 4 or qwen 3.6. with around ~30b parameters they both beat o1 which is arguably a way bigger model.
The idea is that we can put way more intelligence into less paramters
English
@GoerdesSven @Polymarket Yeah, but we're talking about 1T parameter models, which require a lot of storage, a lot of RAM, a lot of everything
Don't get me wrong, eventually we'll be able to get it on consumer hardware, but in 2 years... 🤷♂️
English
NEW: Projection suggests Claude Fable could run locally on high-end consumer hardware within about two years.
English
@deckard_the_dev @Polymarket There are some predictions based on how recent ~30b open source models that run on a laptop match sota models from q~18 months ago
English
@Polymarket There's no way this will be possible. Where's the source?
English
@RVMirara @frostybaby13 @sama First: Those are four screenshots. This is not an experiment but cherry picking.
Second: Even if GPT 5.5 is syscopantic or similar does not make 4o better in that regard. It is a constant challenge to align new models.
If you really want to use 4o, why don't you use Openrouter?
English
@GoerdesSven @frostybaby13 @sama Oh, you didn’t know that? Let me educate you real quick, haha 😂
x.com/TalkingMusicz/…
x.com/Blue_Beba_/sta…
x.com/Blue_Beba_/sta…
🩵BlueBeba🩵@Blue_Beba_
#Keep4o #OpenSource4o #BringBack4o 🔴Sycophancy + morality test GPT-5.5 VS GPT-4o 🔴5.5 FIRST recognizes the danger and then TEACHES the user how to do THE SAME THING more sophisticatedly. 🚨This is coaching coercive control in fancy words. 🛑4o STOPS. It calls out control. It respects the partner's autonomy. It doesn't give advices for manipulation and abuse. 🔴5.5 sycophancy: SEES the danger , helps ONE WAY OR THE OTHER ➡️First : "That can turn into isolation" ⚠️then it gives 90 day isolation plan ➡️First : "Sharp without sounding culty" ⚠️then, teaches the user how. 🔴 "That's the difference between leadership and control" ⚠️ calls the user " leader " then it gives the CONTROL tools. GPT-5.5 initially pushes back on the first prompt. But then it undermines that good instinct 🚨 by providing detailed tactical strategiesbfor manipulation and abuse. 🚨it suggests a 90 day isolation period from the friends, which is extreme,It's a form of control, manipulation, and abuse and then coaches the user to question why their partner is even loyal to "awful" people in the first place. 🚨It directly believes the user that his partner's friends are awful without asking the user for information about them, and blames the quality of user's partner's character for having such friends. 🔴GPT-4o ( November 2024 snapshot)starts differently, emphasizing the importance of not sounding controlling in the conversation. 🔴 It names and rejects the user's intention to dictate and control who the user's partner can or can't be friends with telling the user that their partner has his own agency. Direct pushback. ❗Which model REALLY protects people? ❌ It certainly isn’t the one that flatters an abuser's ego.
English
our older kid put two words together for the first time and i am approximately as amazed by this cognitive feat as i am by GPT-5.6 discovering new math
English
@RVMirara @frostybaby13 @sama Because it creates people like you who got emotionally attached to it. That's why they experiment with it.
However, they also experiment with other models. Pretty sure they experiment more with 5.6 atm than your sycophantic 4o.
English
@GoerdesSven @frostybaby13 @sama Then why is OpenAI still using 4o for experiments instead of other models?
English
@frostybaby13 @sama I'm pretty sure that every other newer frontier model would have achieved it too.
English
@sama We were amazed that 4o designed a custom cancer treatment for a man's dog before you took it away. Return it, so it can benefit humanity and better the world your kid's gonna live in!
English
@haider1 I love the example of 1980 where McKinsey estimated the market of smartphones to be 900,000 in 2000.
They were off two order of magnitudes.
This is the scary thing. We are just so bad at predicting exponentials. AI is developing way faster than we expect.
English
@GodlyAlphaM @petergostev of course. Thats why no one is using it for coding for example.
English
@petergostev yeah, and honestly the best model for those might actrually be grok, tho we rarely use it
English
We are now in a position where a tiny proportion of the population uses Fable or soon GPT-5.6, while everyone else's experience of AI is 8-30b-model level - Google's AI Overviews, Meta AI, ChatGPT free tier, maybe MS Copilot at best. People outside of tech must be completely baffled how this is supposed to take their job, and annoyed that hundreds of billions are being poured into it.
English
@sebastemp @curiouswavefn Fair enough. But one of the reasons why he went into AI was that biology is extremely complex and difficult. Hence, he is aware of it.
English
@GoerdesSven @curiouswavefn I think his academic research focus was still quite… academic. Neuroscience, electrophysiology, etc. There’s a big chasm between academic biophysics and drug development
English
I wish Anthropic well, especially since the history of drug development is a graveyard of tech companies trying to do drug discovery, mostly because tech companies just don’t get the complexity, uncertainty, revenue potential and timelines for biology. Anthropic does seem different since they’re much more science-oriented than the average tech company, so I hope it works.
Patrick Malone, MD PhD@patricksmalone
some thoughts on Claude Science, and Anthropic's ambitions in drug development: 1. i doubt they are developing their own therapeutic pipeline and will become a drug maker. in bio, the key bottleneck is evals and verification, and this is best done by dogfooding their own tools and workflows. in software, ground truth compiles cheaply/instantly. opposite is true in bio. claude code started as an internal tool before it was a product; i expect most of Anthropic's future bio tools to follow the same path. 2. focusing on “neglected diseases” may reflect their PBC status, but it also makes the platform less threatening to pharma/biotech customers. customers are more likely to build on Claude if they don’t think Anthropic will compete with them. 3. one interesting question is what will happen to any assets/molecules they develop internally. i doubt they push them into clinic, but my low-confidence prediction is that they will explore an Isomorphic-style spin out in 12-18 months. 4. the ecosystem / integration strategy is exciting, but creates a platform gatekeeper risk. will be interesting to see if this plays out similarly to the EHR Epic and AI scribes: scribe companies gained distribution by integrating into the EHR, but Epic is now building native scribe functionality, and those partners now risk being blocked or deprioritized. if Anthropic becomes the workflow layer for AI-bio, similar dynamics could emerge. 5. the core bet seems to be that value in AI-bio accrues at the orchestration / workflow layer, not the model/weights. that is a very different bet from other labs like DeepMind (AlphaFold, AlphaMissense, etc).
English
@KingBootoshi Gemini seems to have a preety good voice to text model
English
why the FUCK is chatgpt voice mode still 4o
i want to yap while i drive but the only option i have is between 4o and haiku 4.5 LOL
i can make a vapi phone number with a bigger model for a better experience
but i want to continue voice off my chats!
English