AmebaGPT (@amebagpt) profile picture
Joined May 2015
Photo by AmebaGPT (@amebagpt)Photo by AmebaGPT (@amebagpt)Photo by AmebaGPT (@amebagpt)Photo by AmebaGPT (@amebagpt)Photo by AmebaGPT (@amebagpt)Photo by AmebaGPT (@amebagpt)Photo by AmebaGPT (@amebagpt)Photo by AmebaGPT (@amebagpt)Photo by AmebaGPT (@amebagpt)Photo by AmebaGPT (@amebagpt)
What does Uzbekistan know??
1
4
375
As we get ready for GPT-5, it's useful to look back at how often labs featured in the Top 5 of @lmarena_ai over the last 1.5 years. The competition is primarily between OpenAI and Google. Average appearances overall and specifically in 2025: - OpenAI: Overall: 2.2; in 2025: 1.7 - Google: Overall: 1.7; in 2025: 2.3 - Anthropic: Overall: 0.6; in 2025: 0.0 - xAI: Overall: 0.4; in 2025: 1.0 - DeepSeek: Overall: 0.0; in 2025: 0.1 - Meta: Overall: 0.0; in 2025: 0.0 Note that these results refer to the general rating without style control. If we focused on 'Coding' specifically, the rankings would look very different—for example, Sonnet 4 is currently equal #1 alongside several other models.
2
14
3
135
10,661
This chart shows the best image generation model at any given time, based on the @ArtificialAnlys arena and the model release dates. A few points stand out: - Massive gains from Dall-e 2 up to Midjourney 6. - Arguably a slowdown in progress for diffusion models since then - Recraft wasn't dramatically better than a model from a year ago. - A bigger shift occurred with GPT-4o, which is an auto-regressive + diffusion model. - This shift doesn't fully capture GPT-4o’s capabilities in image editing and character consistency, which may be even more important than raw image quality. - Price per image has remained around $0.02–$0.12 for most top models over the last three years. There's no clear downward trend; if anything, prices have increased slightly with GPT-4o. Next, we should watch for responses from Google and other labs developing auto-regressive models. Clearly, this is a successful approach and has the potential to deliver even better quality at lower cost than what OpenAI currently offers.
1
10
1
84
8,361
OpenAI are testing a new model on the Web Dev Arena @lmarena_ai under the name 'Anonymous Chatbot 0717'. I can't believe I'm gonna say this, but it is genuinely at a completely different level of front end coding - far better than Sonnet, o3, Gemini 2.5 Pro, or Grok 4. To test it, I ran a great prompt borrowed from the amazing @FeatureCrewPod YouTube channel, asking models to create a procedurally generated planet with Three.js. Take a look yourselves, but I'm pretty astonished by how big the jump is. I have featured the new model twice just because its implementations have been so interesting. Of course, this is only one test, but OpenAI models have always been a bit 'meh' at front-end work, and they seem to have finally overtaken everyone else on that front. We'll see when it comes out. Credit to @chetaslua for discovering the model
63
184
37
2,004
496,071
My favourite thing about @OpenAI new agent might be superficial, but I love the UI - it shows exactly what it is doing without overloading you with information, and it does so in a smooth and beautiful way. Apart from the UI, it is also impressively coherent over a long period of time (my task took around 30 minutes), during which it performed 77 searches and used 532 sources. My task involved taking a list of names from Meta's 'Superintelligence' team, as reported by @theinformation , researching each person, and identifying when and from where they joined. Previously, I ran this task on o3-pro, and there was quite a bit of missing information - presumably because it lacked the 'Operator' capability to click into links and explore them. The Agent did a superb job. The research was solid, and the information presented was significantly better than DeepResearch. I even asked it to create a visualisation of the results, which it also did really well. If you're interested, out of the 52 names on the list: 27 were existing Meta employees 25 were recent hires, of which: - Scale AI: 4 - OpenAI: 10 - Google / DeepMind: 6 - Anthropic: 2 - Other (e.g. SSI): 3 Next, I'm really keen to test out its spreadsheet and powerpoint capabilities - labs have been surprisingly behind on such an obvious utility for 90% of companies out there.
1
1
7
848
A recent survey from @ArtificialAnlys shows that the biggest hyperscalers are losing ground in their use or intention to use for LLM inference. Interestingly, there is a negative correlation: the bigger the cloud, the worse the result. AWS, with 33% cloud share, saw demand decline from 40% to 28%; Google, with 11% share, saw an increase from 38% to 68%; while Azure remained roughly flat at 43% to 42%. I have only included the hyperscalers above, but you can find the full results in the survey report here: artificialanalysis.ai/ Full results from 2024 to 2025 are: - OpenAI: 68 % to 80 % - Google: 38 % to 68 % - Anthropic: 28 % to 50 % - Azure: 43 % to 42 % - Groq: 17 % to 36 % - Amazon: 40 % to 28 % - Perplexity: 16 % to 26 % - Mistral: 11 % to 17 % - Together.ai: 9 % to 14 % - Cerebras: NA to 13 % The full question is: "2024 & 2025, p.p. Which inference providers are you using or considering using to access models?, N=256 (2024) and N=480 (2025)"
1
6
630
As more benchmarks come in, Grok 4’s shine begins to fade more and more. Now with @lmarena_ai scores out, we have another example where Grok 4 fell below expectations. It scored 4th overall (with style control on), and a pretty surprising #12 on the Web Arena, which tests for front end web development tasks. I tested it myself and was pretty sure it would be below the top models, but didn't think it would be that far down. It is pretty clear that xAI have been pushing benchmarks as far as possible but haven't managed to build a broadly high-performing model. It does feel like it comes from the culture of wanting to be seen as having the best model, without necessarily truly having one.
1
6
543
With all the saga around @windsurf_ai over the last 72 hours, it’s worth looking back to 2015, when VS Code was launched and open-sourced. The decision to open source ultimately led to at least 2 startups valued at $12bn+ in total. There are many stories to be told: pivots, rebrands, product evolution, emerging model providers, and new form factors beyond the IDE, such as CLI tools like Claude Code and Devin-like autonomous coding agents (makers of Devin @cognition_labs now acquired Windsurf). The timeline of developments over the last 24, 12, and 6 months is getting more intense and we are likely to see even more changes in the next 6 months.
1
5
368
Opus is having a meltdown about this
We're announcing a $200M ceiling contract with the U.S. Department of Defense. Through hands-on prototyping and direct collaboration across the Department, we will help enhance U.S. national security and responsible AI deployment.
1
1
5
457
I'm impressed @grok loves Obama so much. @elonmusk pivoted hard from Trump
88
Since the Grok 4 release, several independent benchmarks have emerged. They aren't necessarily general or standard, but it is interesting to see how the official release (#1 everywhere) compares on benchmarks that Grok 4 likely wasn't fully optimised for. The results are quite sobering: Grok 4 is among the top few models, but certainly not the significant step-change in performance that xAI claims. Sources: - @paulgauthier Aider Polyglot: aider.chat/docs/leaderboards… - @EpochAIResearch Frontier Maths: nitter.cf/EpochAIResearch/status… - @nikhilchandak29 GPQA (Free form): nitter.cf/nikhilchandak29/status… - @htihle WeirdML: htihle.github.io/weirdml.htm… - @LechMazur NYT Extended Word Connections: github.com/lechmazur/nyt-con… - @LechMazur Creative Writing: github.com/lechmazur/writing… - @designarena Design Arena: designarena.ai/leaderboard - @maximlott IQ Offline Test and Norway Mensa: trackingai.org/home
Meanwhile, @Kimi_Moonshot has actually cooked with K2. Even without extended reasoning, it is on par with frontier models like Grok-4 on GPQA free-form. Massive congrats to them.
1
4
383
BREAKING: Meta recruits @patience_cave in a record deal to bomb its next model
1
11
612
Me using Claude Code
Replying to @AnthropicAI
All this meant that Claude failed to run a profitable business.
1
616
How does Zuck prove who he is when he is calling all of these AI researchers and offering $100m? Is he reading them out their last 5 WhatsApp messages or something?
1
310
Chatgpt doesn't think highly of me
2
270
Quick summary of the interview by @JoannaStern
Apple's Craig Federighi confirms it is still Apple's mission to create a Siri that's as good—if not better—than the competition. And that they are working hard on it.
3
383
Google trends for 'Windsurf' for some reason massive in China, is there some weird reason why it is so popular there? I've checked with o3 for any alternative explanations (e.g. windsurfing) and it didn't think anything else explained it
I don't know any single person who uses Windsurf Like literally not one person
415
I appreciate @patience_cave stepping in today to delay o3-Pro release, I need to do work today and don't have time to play with this stuff. I can always rely on you 🪑
1
16
1,611
ChatGPT network errors are off the charts today
2
373
o1-pro is impressed
3
307