Basil Frankweiler

5.9K posts

Basil Frankweiler

Basil Frankweiler

@BasilFranken

Beigetreten Nisan 2024
533 Folgt74 Follower
Basil Frankweiler
Basil Frankweiler@BasilFranken·
@eduleadership @pseudonym_witty @hecubian_devil Education Philosopher? Holy freak. Dude, get a real life. You know nothing about hyper-gifted children. Put a kid who tests at the masters level at 9 years old in a class with most teachers and see what happens.
English
1
0
0
17
Justin Baeder, PhD
Justin Baeder, PhD@eduleadership·
@pseudonym_witty @hecubian_devil Educators should specifically stifle disagreeableness, and cultivate positive personality traits. Yes, people can be successful if they are low in agreeableness, but it’s still a positive trait. Don’t be fooled by survivorship bias.
English
2
0
7
345
Cassie Pritchard
Cassie Pritchard@hecubian_devil·
I think you were actually just dumb and ill-behaved, tbh. I don’t think actually-smart kids experienced this. They were generally liked and pampered by teachers.
memetic_sisyphus@memeticsisyphus

It’s true, every halfway intelligent right winger I know irl had a massive conflict with at least one elementary teacher over things like: reading ahead, reading too difficult books, not showing enough work, etc etc. it’s the first time we experience the uncaring tyranny of state bureaucracy and it sucks.

English
407
151
3.7K
378K
Basil Frankweiler
Basil Frankweiler@BasilFranken·
@eduleadership @pseudonym_witty @hecubian_devil I won't share my background but I can tell you with 100% certainly that MOST teachers love extremely gifted kids. Many do not. This really should be self-evident to any adult. Self-referencing PhDs excluded I guess.
English
0
0
0
5
Justin Baeder, PhD
Justin Baeder, PhD@eduleadership·
@hecubian_devil I think a lot of adults tend to reframe their poor childhood behavior as suffering at the hands of cruel teachers who didn’t appreciate their genius—it’s a better story. Your teacher wouldn’t let you read a book? Come on. You got in trouble for mouthing off, not reading.
English
26
13
354
41.7K
Padder Stoel
Padder Stoel@Padderstoel·
@pepel_klaasa Trump's inept and irresponsible military strategy is costing the US in credibility. Russia and China both know that the US does not have the stocks for a prolonged conflict anymore. And Trump does not care because he believes he wont be first in the firing line
English
6
0
10
533
Pepel Klaasa
Pepel Klaasa@pepel_klaasa·
The United States is holding up virtually all weapons deliveries purchased by EU countries. It is not like we’re completely defenseless here, but it’s obvious that if Russia attacks within the next year or two, we can’t rely not only on U.S. assistance – but even on weapons supplies. For Estonia, this is especially critical. news.err.ee/1610000377/est…
English
535
2K
5.6K
414.6K
sammy the stubblejumper
sammy the stubblejumper@steveofthewater·
@pepel_klaasa @MooseandRaven Fuck America. Time for Europe to start making guns & artillery shells and getting into rapid drone research and development, as if they stand alone against the whole rest of the world. If America falls to fascism they will certainly align with other dictators against democracy.
English
18
0
1
756
Basil Frankweiler
Basil Frankweiler@BasilFranken·
@AGORACOM @WesRoth If I could have caught him walking out or something I would have. I figured pulling up a chair at his table was a bit much. lol
English
1
0
0
2
Wes Roth
Wes Roth@WesRoth·
Newly released web traffic data reveals a significant shift in the consumer generative AI landscape. OpenAI's ChatGPT has seen its commanding market share erode, dropping from a dominant 77.43% down to 56.72%. As OpenAI's dominance wanes, its primary competitors are rapidly accelerating. Google's Gemini has surged from a mere 6% share to capture 25.46% of the market, while Anthropic's Claude has more than quadrupled its footprint, growing from 1.4% to 6.02%.
Wes Roth tweet media
Similarweb@Similarweb

Gen AI traffic share update Main takeaways: → Gemini holds a quarter of the share. → Claude almost doubled its share between February and March, crossing the 6% mark. → DeepSeek surpassed Grok again. → We’ve added m365.cloud.microsoft/chat to Copilot’s share, which explains the increase from previous updates. 🗓️ 12 months ago: ChatGPT: 77.43% Grok: 7.03% Gemini: 6.00% DeepSeek: 3.73% Perplexity: 1.66% Claude: 1.40% Copilot: 1.38% 🗓️ 6 months ago: ChatGPT: 71.75% Gemini: 13.56% Grok: 4.08% DeepSeek: 2.81% Perplexity: 2.51% Claude: 2.26% Copilot: 2.01% 🗓️ 3 months ago: ChatGPT: 63.19% Gemini: 22.59% DeepSeek: 4.08% Grok: 3.26% Claude: 2.22% Perplexity: 1.86% Copilot: 1.80% 🗓️ 1 month ago: ChatGPT: 56.72% Gemini: 25.46% Claude: 6.02% DeepSeek: 3.74% Grok: 3.44% Copilot: 1.99% Perplexity: 1.64%

English
13
4
65
9.7K
Basil Frankweiler
Basil Frankweiler@BasilFranken·
@WesRoth Synthetic data is gonna blow the lid off AI. You heard it here first.
English
0
0
0
2
Wes Roth
Wes Roth@WesRoth·
A joint team from CMU Robotics and Lambda has introduced "Sim2Reason," a breakthrough training methodology that teaches LLMs physics by immersing them in virtual simulators rather than relying on human-written textbooks. The bottleneck in STEM AI is the scarcity of human-annotated data. The team bypassed this entirely by procedurally generating millions of random, physically grounded scenes inside the MuJoCo physics engine using a custom domain-specific language (DSL). They auto-generated verified Q&A pairs and RL-trained the models purely on this synthetic data, with zero human labeling in the loop. The zero-shot performance gains are massive. By learning the physical meaning of equations through simulation, models showed across-the-board improvements, including a +7.5% jump on the International Physics Olympiad (IPhO) for Qwen2.5-3B, outperforming models trained on thousands of curated, real-world physics problems.
Mihir Prabhudesai@mihirp98

What if AI learned physics the way Newton did – by experiencing it? We built Sim2Reason: train LLMs inside virtual worlds governed by real physics laws, zero human annotation. Result: +5–10% improvement on International Physics Olympiad, zero-shot. 🧵

English
4
6
34
4.5K
Basil Frankweiler
Basil Frankweiler@BasilFranken·
@AGORACOM @WesRoth In a local eatery last night and I overheard a bricklayer talking about how great Claude was. I was dying to go talk to him but refrained.
English
1
0
0
5
AGORACOM - George
AGORACOM - George@AGORACOM·
@WesRoth Did they "lose" it ... or did a bunch of Google users who never used AI start engaging with Gemini .... users OpenAI never would have got anyways
English
1
0
0
22
Basil Frankweiler
Basil Frankweiler@BasilFranken·
@lukeNukemAI @WesRoth Good reason to think it is not. Opus 4.6 found like 500 zero days in github. It invented a new way to look for bugs. It looked at the commits looking for clues. It was not told to do that. So add 40 or 50% more ability to that and I can see it being a problem.
English
0
0
0
9
JM
JM@lukeNukemAI·
@WesRoth And you don’t think that is hype?
English
1
0
1
71
Wes Roth
Wes Roth@WesRoth·
A new report from Bloomberg details a critical, alarming moment inside Anthropic when researchers discovered that their highly anticipated, unreleased "Mythos" AI model successfully and autonomously breached simulated computer networks within hours during internal red-team testing. Moving far beyond standard code generation, the Mythos model reportedly demonstrated advanced, unprompted offensive cybersecurity capabilities. It exhibited the ability to actively probe for vulnerabilities, exploit zero-day flaws, and execute lateral movement through secure environments without human guidance.
Wes Roth tweet media
English
12
10
53
3.8K
Basil Frankweiler
Basil Frankweiler@BasilFranken·
@SayItLoud19 @WesRoth Actually Qwen3.6 figured this out on the first turn last night. (second chat, first chat it said walk, so I asked it how I was going to wash the car if it was 100 meters away, so 2 turns.) Gemma 4 can solve it but it recognized it and told me that it had gone viral.
English
0
0
0
29
A fierce pancake
A fierce pancake@SayItLoud19·
@WesRoth Enough with this 'scary' clickbait bollocks. It still can't work out you need to take your car to a car wash to get it washed. It has the consistency of a three year old.
English
1
0
4
100
Basil Frankweiler
Basil Frankweiler@BasilFranken·
@WesRoth 3.6 MUCH faster than 3.5. Ironically 3.5 27B is still the coding benchmark leader but 3.6 35B is neck and neck with it but faster. Of course more memory. You can squeeze 27B in a 24GB machine. Gemma 4 is still the best overall, tho not the best coder.
English
0
0
0
11
Wes Roth
Wes Roth@WesRoth·
Alibaba open-sourced Qwen3.6-35B-A3B under a permissive Apache 2.0 license. It is a highly efficient sparse Mixture-of-Experts (MoE) model that contains 35 billion total parameters but only activates roughly 3 billion during inference. Despite its incredibly lightweight active footprint, the model delivers agentic coding and reasoning performance on par with models ten times its active size, soundly outperforming dense predecessors like the Qwen3.5-27B.
Wes Roth tweet media
Qwen@Alibaba_Qwen

⚡ Meet Qwen3.6-35B-A3B：Now Open-Source！🚀🚀 A sparse MoE model, 35B total params, 3B active. Apache 2.0 license. 🔥 Agentic coding on par with models 10x its active size 📷 Strong multimodal perception and reasoning ability 🧠 Multimodal thinking + non-thinking modes Efficient. Powerful. Versatile. Try it now👇 Blog：qwen.ai/blog?id=qwen3.… Qwen Studio：chat.qwen.ai HuggingFace：huggingface.co/Qwen/Qwen3.6-3… ModelScope：modelscope.cn/models/Qwen/Qw… API（‘Qwen3.6-Flash’ on Model Studio）：Coming soon～ Stay tuned

English
7
4
55
3.9K
U.S. Navy
U.S. Navy@USNavy·
First contact. 🛸👽

On April 10, U.S. Navy divers were the first on the scene as the Navy and NASA successfully recovered the Orion spacecraft and its crew in the Pacific Ocean, bringing four astronauts safely back to Earth after their journey beyond the Moon.
English
337
2.3K
14.8K
452.2K
Basil Frankweiler
Basil Frankweiler@BasilFranken·
@WesRoth You didn't say the word "Insane" anywhere. They're gonna take away your youtube card.
English
1
0
1
19
Wes Roth
Wes Roth@WesRoth·
Anthropic rolled out Claude Opus 4.7 across its native platform and major cloud providers, marking its most capable frontier model to date with a heavy emphasis on autonomous, long-running execution. The update introduces a massive vision upgrade, tripling the native image resolution capacity to over 3.75 megapixels. This sharper visual processing directly enables the model to output significantly higher-quality UIs, slide decks, and complex formatted documents. To manage this advanced reasoning, Anthropic added a new xhigh effort level in the API for granular latency control, beta Task Budgets to manage costs on extended runs, and a dedicated /ultrareview command in Claude Code to act as a meticulous, automated code reviewer.
Wes Roth tweet media
Claude@claudeai

Introducing Claude Opus 4.7, our most capable Opus model yet. It handles long-running tasks with more rigor, follows instructions more precisely, and verifies its own outputs before reporting back. You can hand off your hardest work with less supervision.

English
3
5
28
2.6K
Craig Hewitt
Craig Hewitt@TheCraigHewitt·
Ok, time to retire this beauty. what should i get?
Craig Hewitt tweet media
English
11
0
9
2.1K
Basil Frankweiler
Basil Frankweiler@BasilFranken·
@cbames If I could wave a wand and get rid of a blocker, I'd probably pick motor controllers. (I mean... battery tech sure but that's a given) If we could switch DC loads with 1/10 the cost in size, money and heat we'd have something.
English
0
0
0
19
Barrett Ames
Barrett Ames@cbames·
@BasilFranken The simplest thing to point to is the hardware software divide. It's actually always been both, some algorithms work be with certain hardware.
English
1
0
2
35
Barrett Ames
Barrett Ames@cbames·
Few understand how much false dichotomy holds back robotics.
English
2
0
10
725

Entdecken

@eduleadership @pseudonym_witty @hecubian_devil @Padderstoel @pepel_klaasa @steveofthewater @MooseandRaven @IrishPhinFan