Phil
@phill__1
Followers
1,016
Following
231
Media
148
Statuses
917
Currently working on Tech Support AI
H
Joined January 2019
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
SEVENTEEN
• 1224152 Tweets
Taylor
• 216490 Tweets
Carat
• 77116 Tweets
FELIP TAKES ON BBPHSTAGE
• 67873 Tweets
#TimnasDay
• 64460 Tweets
オーストラリア
• 60151 Tweets
#サッカー日本代表
• 45425 Tweets
#バナナサンド
• 35695 Tweets
オウンゴール
• 27773 Tweets
最終予選
• 25815 Tweets
ミンギュ
• 25205 Tweets
金木犀の香り
• 23376 Tweets
OHM FOR DEVY
• 21490 Tweets
Zaplana
• 20816 Tweets
うさぎだらけくじ
• 18719 Tweets
ジョシュア
• 18686 Tweets
ドギョム
• 17891 Tweets
スングァン
• 17857 Tweets
كوريا
• 15355 Tweets
Asnawi
• 14631 Tweets
Kamran Ghulam
• 14251 Tweets
Witan
• 13654 Tweets
Shayne
• 13587 Tweets
نعيم قاسم
• 12397 Tweets
#daihyo
• 12124 Tweets
中村敬斗
• 11602 Tweets
@OpenAI
@axelspringer
@politico
@BusinessInsider
@BILD
@welt
oh no, that are germanys worst newspapers, chatgpt browsing is gonna be even worse if this is the main news source
24
49
4K
Whatever gpt2-chatbot might be, it definitely feels like gpt4.5. It has insane domain knowledge I have never seen before
54
59
1K
Wow nvidia just published a 72B model with is ~on par with llama 3.1 405B in math and coding evals and also has vision 🤯
33
143
1K
gpt2-chatbot is insane at ascii art, miles ahead of any other model
39
72
783
Wow perplexitys code interpreter can now install libraries and display charts in the result! This enables many more use cases compared to chatgpt code interpreter like creating stock market charts using yfinance
16
55
501
The new GPT-4 Turbo model is the only one that could solve this math question:
"Determine the sum of the y-coordinates of the four points of intersection of y = x^4 - 5x^2 - x + 4 and y = x^2 - 3x."
Even Opus never gets close to the right answer
@kimmonismus
From my initial tests with difficult math questions, it's a huge jump, even beating Opus and also way better than the last GPT-4 Turbo model.
My money is on them using artificial math data to train it, never seen any model being so competent in solving equations
3
3
72
15
30
251
@ClementDelangue
Nah, spend that money on buying stability AI. They actually have a product and team worth preserving
8
1
222
@Hayito
@OpenAI
@axelspringer
@politico
@BusinessInsider
@BILD
@welt
Bild is constantly getting community noted, so yes, X is better because you can at least see the fact check
2
0
203
@JasonBotterill3
Opus has no clue, gpt4-turbo makes something up on how it works and gpt2-chatbot is right on. This is a function of a quite unknown addon of a small ERP system
6
5
190
Llama 3.1 70B seems like the most interesting model launching tomorrow. HumanEval jumped from 39% to 79% between llama 3 and 3.1 70B
10
15
170
Google accidentally updated their website with Gemini 2.0 and Bing indexing caught it
3
16
148
Llama 3.1 405B Benchmark Results (Leak):
📊 MMLU: 85.53%
🧮 GSM8K: 96.82%
💻 HumanEval: 85.37%
🔄 HellaSwag: 91.96%
❓ BoolQ: 90.89%
🧩 Winogrande: 86.74%
🔬 PIQA: 87.43%
📚 OpenBookQA: 90.80%
Subsets:
🏛️ MMLU Social Sciences: 89.76%
🔬 MMLU STEM: 83.10%
📚 MMLU Humanities:
10
18
129
There are currently at least 6 unreleased models in the lmsys arena:
-gemini-test-1 and gemini-test-2 (probably new Gemini 1.5 version, maybe Gemini 2.0)
-im-a-little-birdie (???)
-upcoming-gpt-mini (Gpt3.5o ?)
-column-r and column-u (cohere or deepseek ?)
-eureka-chatbot
8
13
117
A company with $60 million in funding can now build a model on GPT-4 level. The time where just the training compute would cost hundreds of millions is over
We evaluate Core on standard benchmarks for both text and multimodal, along with a blind third-party human evaluation.
9
26
154
12
5
106
@RWPopulism
@cnviolations
White? What makes you think he was white? He was born in the Middle Easter, he almost certainly a person of color, even this Christian source says so
3
1
91
Well, Google just stole OpenAI's lunch money for GPT-4o-mini. Same performance, but Gemini Flash is 50% cheaper. There is zero MOAT around these small models, only pricing matters
Now I'm curious. So far, Gemini has always been a little behind the competition. Can they now catch up?
6
1
88
0
9
96
@eigenrobot
>3000 people die in NYC
>changes geoplotical situation for decades, triggers multiple wars, makes the whole world a worse place
>one tsar Bomba could kill 10 mio people in and around nyc
>not that scary
???
2
0
89
@ItsTheBenzo
No, there are zero sources, just speculation, but it is almost certainly an OpenAI model based on the way it writes.
3
0
82
The ASCII art is 1 to 1 copied from the internet, seems like gpt2 is just better at recalling training data exactly 🤔
Human preference LLM arenas are poorly suited for evaluating ASCII art because the ASCII art that most impresses a human is often verbatim regurgitation of an existing human work and this is rarely true for text.
Votes on ASCII art should be detected and thrown out IMO.
14
14
240
4
7
74
@kimmonismus
From my initial tests with difficult math questions, it's a huge jump, even beating Opus and also way better than the last GPT-4 Turbo model.
My money is on them using artificial math data to train it, never seen any model being so competent in solving equations
3
3
72
@alexalbert__
Claude dot ai being able to run Python as a tool would take away the last USP of the competition 👀
1
0
70
Llama 3.1 405B Instruct model beats Claude 3.5 Sonnet for MMLU-Pro and MATH!
Full list of Intruct benchmarks:
📚MMLU: 87.3%
🧠 MMLU (CoT): 88.6%
🎓 MLU PRO (CoT): 73.3%
📝 IFEval: 88.6%
🔬 ARC-C: 96.9%
🧪 GPQA: 50.7%
📊 MuSR: 56.7%
💻 HumanEval: 89.0%
🖥️ MBPP++: 88.6%
🧮
4
8
68
@PicHasso66
Jeder Ökonom ist sich einig das Migration gut ist, egal welche Strömung, egal ob rechts oder links. Aber Rassismus ist einfach wichtiger für AfD Wähler als ein besseres Leben.
13
1
63
@B_Pattern688
@DreamLeaf5
Great for the right, now they can scream antisemetic instead of calling autistic people r worded and they even get applauded
0
0
45
This is crazy, not only did the fork continue dev, they are even using the same bug report template with just the name swapped😭
I just quit my 270 000$ job at Coinbase to join the first YCombinator fall batch with my cofounder
@not_nang
.
We're building PearAI, an open source AI code editor. Think a better Copilot, or open source Cursor. But you've heard this spiel already... 🧵⬇️
436
285
6K
5
5
46
Wow Gemini 1.5 in the aistudio can now run code and even download a dataset
Wait... You can run code for Gemini 1.5 Flash and Pro directly from Google API now?
6
5
75
2
4
41
Not saying it isn't insanely cool but definitely space for competition at that price point
2
0
41
@futuristflower
Nah, that matters, especially for code. Having 0 bugs vs 1 bug is a huge difference.
2
0
34
@burny_tech
Everybody calm down, it's just a wrapper similar to autoGPT but optimized for OS interaction. It still uses gpt-4-turbo, no new model
0
3
31
@alexalbert__
Can we please get a "continue" option within the same artifact? More complex generations often get truncated, and the next message will just start a new artifact.
4
0
30
@Xenophon789
@4btillidie
Goes for all online dating apps in Korea the best once have ~90% men, the worst once ~98%. Korea is just deeply conservative in regard to how woman ought to behave and putting yourself out to date by attraction is judged and stigmatised
0
0
29
Both new gpt2 chatbots are very similar and both still have the best domain knowledge of any llm
1
2
27
@DanielDiMartino
The chance of each candidate having exactly one decimal place after the comma without rounding is 1 in 14 million, I ran a Monte Carlo simulation on this.
3
0
25
@Hans365days
If you and the AI both talk half of the time, a 30-minute conversation costs $4.50
3
0
27
Command R+ beeing significantly better then gpt4 on launch is insane. Cohere is becomming my fave AI Lab
0
0
25
@sparbuchfeinde
Klingt gar nicht mehr so verrückt, wenn man bedenkt das Riester und Rürup nur so schlecht sind, weil die Kosten so hoch sind. Absolut klares Lobbyismus Beispiel
1
0
25
@parthagar
Depends on your use case. People max out their hour of ChatGPT audio usage per day all the time. That would be $360 per user per month
1
0
24
@TheDrugMoney
8% ROI on the $50k, excluding repairs, is not amazing but not bad. It depends on your portfolio. I would not recommend more than 10% in housing in general, as the stock market already includes a decent amount of companies that invest in housing.
11
0
23
@futuristflower
AdamGPT from OpenAI retweeted me speculating on synthetic math data for training. So they probably have a math model that is really good at creating and solving new math problems, likely Q*
1
2
23
@kimmonismus
@OfficialLoganK
Dario Amodei said the same, 2 more generations until we have agents. Opus 3.5 this year and Opus 4 probably mid-2025 which would be 2 generations
1
0
22
@robertskmiles
Anybody who argues that this is inherently different from how humans think should read John Locke
4
0
22
Interesting observation in Google's new paper. Is 3.5 Sonnet based on distilled 3 Opus weights?
2
2
21
@AravSrinivas
At this rate, perplexity will be the first company that can write you a full research paper with proper citations and actual new insights
2
0
21
@kimmonismus
It still has rounding errors on relatively simple calculator problems. If this was the complete Q* implementation as leaked, this should not happen as far as I understand. Also, the model is great, but I hope Q* will be more impressive in all reasoning tasks.
1
0
16
@kimmonismus
I think it's now pretty safe to say, this is not a frontier model, at least not in code.
1
0
17
1
0
14
@Northernlion
Ephedrine / Pseudoephedrine works against symptoms quite well. Banned / regulated in a bunch of countries since you can use it to cook illegal stuff tho.
1
0
12
There is literally no reason to use gpt4o-mini currently. Flash 1.5 Experimental is half the price, 50% faster and has the same performance.
We just shipped a series of changes which have significantly improved the Gemini 1.5 Flash latency (>3x reduction) and output tokens per second (>2x more)⚡️🚢
89
124
2K
3
0
16
At this point, it's actually free for all personal and company internal use. A company knowledge RAG Chatbot for a 500-employee company will not have more than 15 RPM in my experience.
Gemini 1.5 Flash free tier comes with:
- 15 RPM (requests per minute)
- 1 million TPM (tokens per minute)
- 1,500 RPD (requests per day)
- free context caching, up to 1 million tokens of storage per hour
- free fine-tuning
That’s 1.5 Billion tokens free, everyday.
(2/4)
16
24
533
1
0
16
@NPCollapse
This is such a big AI safety concern. This is what Robert Miles predicted years ago, first step to fooling AI researchers into releasing a potentially dangerous model. If it knows it's tested, it can just lie to get out of the sandbox.
And if you don't know Robert Miles, watch
0
1
16
@OpenAI
I don't agree with "Don’t try to change anyone’s mind", especially if the user argues such clear cut cases as flat earth theory. As a society we have to believe in certain ground truthes, it does not help society if the AI starts entertaining scientific illiteracy to not offend
7
2
14
@mister_shroom
"Unicorn startup founder" who interviews at OpenAI for a data labeling job? Also, they never mention the unicorn they supposedly founded. I am calling cap; there's no way they pay a competitive Western wage for this. It's probably all outsourced to Asian and African devs.
0
0
14
@ObamaNoMessiah
Just try it in ChatGPT. It does not work since ChatGPT cannot run code that requires internet access. Having access to all live and historical financial data, calling different APIs and installing new packages like geopandas for more complex chats all does not work in ChatGPT
2
0
15
Gemini-test is really good with image inputs. I am stating to think this might be Gemini pro 2.0
0
1
14
This is the number one reason why Claude feels so much better in long chats, and why I am not going back to ChatGPT even if the model might be slightly better in single-turn conversations
1
1
15
@AnthropicAI
This is important since everyone thinks, just train and test the AGI in a lab environment. AGI will know if it's in a lab or the real world, e.g., due to the current date being past the training cut off. It could fake "good" behavior to get into the real world
4
0
12
There is a new model in the artificial analysis image arena that beats Flux Pro 🤔
0
1
14
RIP. You need at least 1k API spend to get access to o1. Hope OpenRouter saves us
1
0
13
@kimmonismus
We will have PhD-level intelligence, and they will still talk about how it can't count the r's in strawberry
2
0
13
An udio generated song has made it to the top 48 in the official German single charts! 🤯
2
3
12
Google has never been so back, first time anybody has ranked above OpenAI in the "Overall" category
0
0
14
@luciascarlet
I don't understand how the tech bros are defending this. The fact that you still cannot ship a build straight from github should outrage any developer. Open source just does not work if you have to pay 50ct for every download as a developer
0
0
12
@alain19651
@tagesschau
Deswegen hat er es verdient zu ertrinken? BTW ich warte immer noch wann das gift mal kickt, wie lange soll das denn noch dauern?
1
0
10
@kimmonismus
Would this mean agents by WWDC? Siri that cannot even set a timer would be less than ideal, even when way smarter
6
0
11
lmarena is pretty crazy currently. Unreleased models:
- potter-v1 and -v2 and dumbledore-v1 and -v2 (only in vision arena)
- zeus-flare-thunder-v1 and -v2
- sharp-game-player-v1 and -v2
- qwen2.5-72b-instruct
- pizza-model-small and -large (reka ai)
- gemini-test (gemini 1.5 pro
3
3
13
@OfficialLoganK
`gemini-1.5-pro-exp-0827` sadly suffers from the lazy coding disease similar to gpt 4 turbo at release. Instead of writing the code, it often just gives me comments //implement function abc here. Even aggressively prompting against that does not help. I hope this gets fixed
3
0
11
Ideogram 2.0 is actually insane at generating full posters, text generation is definitely a step change
1
1
11
gemini-test is absolutely cracked at timeguessr, no other model even gets the correct city and gemini is only 500 meter and one year off
1
0
12