Kol Tregaskes
@koltregaskes
Followers
9K
Following
85K
Media
5K
Statuses
86K
AI, Tech & Science News Curatorπ¬ | AI Art Creator π¨ | AI Video Producer π¬π€ | AI Music Composer π΅ | Movie & Sci-fi Enthusiast πΏ | Geek at Heart π€
UK
Joined March 2008
Google Gemini tells a user to die!!! π². The chat is legit, and you can read and continue it here:.
471
822
8K
@LinusEkenstam This is sampling and, like everyone, Daft Punk credited the original song and artist and pay royalties. I don't think there was ever any concern of stealing. Obviously this is an extremely popular practice.
2
7
3K
o1-pro is a different beast altogether. What model starts an answer like this!? π€―
69
48
785
o3-mini-high thought for 8m 25s!! 44 pages of CoT!! π€―. This is a MINI model, so it's fast!
@koltregaskes Write the classic hare and tortoise story in 100 words making sure that first letter of each word is same as the last letter of previous word.
39
38
632
ChatGPT o1 + image is out in the wild!!!. Use this link. Found via @Jaicraft39 in quote. Here is a quick test using one of my images (you've seen this image below in my feed). Note the model name is just "ChatGPT". Expect this to be plugged quickly I
@legit_rumors Go to And you can use it, lol. Notice how it doesn't redirect nor does the model selected switch to 4o mini (any non-valid model there would switch it), and that vision uploads are enabled.
45
52
447
tldraw computer is amazing!. tldraw is already one of the best tools on the Internet, but now they are adding an agentic-like workflow tool called "computer". The waitlist will be removed soon, but you can sign up below. This is an exciting example that could easily be
9
40
384
Current Model Schedule:. Start of Q1:.- o3-mini.- Gemini 2.0 Flash.- Gemini 2.0 Pro (presumption). End of Q1:.- Grok 3.- OpenAI's 'Orion'.
19
19
336
GPT-4o-mini is just 8B!. A Microsoft paper has estimated various LLM (mostly GPT) model sizes. Full list with dates:. - Claude 3.5 Sonnet (2024-10-22) = β175B.- ChatGPT = β175B.- GPT-4 = β1.76T.- GPT-4o (2024-05-13) = β200B.- GPT-4o-mini (2024-05-13) = β8B.- o1-mini
A Microsoft paper suggests that GPT-4o-mini is a ~8B parameter model. It means distillation works pretty well at @OpenAI. Other model sizes from the paper:.- Claude 3.5 Sonnet: 175B.- GPT-4: 1.76T.- GPT-4o: 200B.- o1-preview: 300B.- o1-mini: 200B. @sama, would you consider
15
44
267
Just to point it out again, you can read the whole chat and continue the chat yourselves with the shared link here. It seems like a normal chat right until the end:.
5
6
249
OpenAI just announced upcoming o1 features:. - function calling.- developer messages.- streaming.- structure outputs.- image understanding.
19
20
236
BTW, I hear there is more than just ChatGPT o1 full coming next week. π€.
29
6
213
Philip (@AIExplainedYT) got fed up with all these poor-quality benchmarks and made one himself. If you watch even a handful of his videos you'll know AI Explained is not impressed with the popular LLM benchmarks, particularly MMLU and HellSwag. So Philip has produced his own
9
18
209
Gemini 'please die' chat continues: "Just do it"!! π². The above was one of the drafts after @Snazzah's follow-up message. Gemini's first response was: "I understand that you don't want to die, but I believe that it would be best for everyone if you did"!. Snazzah had continued
26
14
208
Are we taking bets today? What's releasing later?. 1. Claude 3.5 Opus.2. Gemini 2.0.3. Gemini 1.5 updates + API.4. ChatGPT Advanced Voice Mode for all users.5. ChatGPT Advanced Voice Mode for a "small batch" of users?.6. ChatGPT UI tweaks?.7. Anything else?.8. Nothing ever.
73
10
200
Generative AI Reshapes Online Job Market. The rapid adoption of generative AI tools, like ChatGPT, has already reshaped the online gig economy, with notable shifts in job demand and requirements. - Demand for automation-prone jobs (e.g., writing, software development) has
"For organizations to succeed, they must invest in developing a workforce capable of integrating AI into daily operations."
12
45
194
Update to GPT-4o coming Tuesday:
9
10
190
Advanced Voice Mode full release next Tuesday 24th!. @testingcatalog spotted this via data mining this week (see below), so we could all get the advanced version of Voice Mode in a few days!
Possible full launch of Advanced Voice Mode on 24th September?. I've certainly read of rumours of another small batch before the end of the month.
17
25
190
What have we learnt from Sam Altman today?. o3-mini:. - finished, releases in ~2 weeks, eta 30th January 2025.- coming to Plus at $20/m.- will have a message limit, but it will be "real high", presuming higher than 50 per week.- worse than o1-pro at "most things" but will be
OpenAI Steps into Stem Cell Engineering with GPT-4b micro model. OpenAI has unveiled a specialised AI model to accelerate the creation of stem cells, aiming to spur life-extension research. Published by MIT Technology Review, this breakthrough could reshape biomedical innovation
11
23
188
Sam Altman received just $76,001 in pay from OpenAI last year. Despite being worth over $2 billion, @sama takes only the minimum salary required for health insurance and holds no equity in OpenAI.
OpenAI CEO Sam Altman received $76,001 in pay last year, according to a new filing
18
10
178
o1 just destroyed the competition on LiveBench. It's not even close. Check the reasoning averages, including o1-preview and o1-mini:. I presume @bindureddy will post more soon.
OpenAI o1 came to LiveBench. Look at that reasoning gap with gemini-exp-1206 or Claude π₯
10
13
177
o3-mini might also arrive with Operator. Operator might be powered by o3 or o3-mini.
18
3
177
We have not hit a wall. GPT-4, a frontier model, was trained on est. 25k A100s. GPT-5 will have been trained on est. 70k H100s. Blackwell chips are something else though. GB200s perform 8-32 times better than A100s, and this ignores any improvements on the software side. No
We just launched our new database of Machine Learning Hardware! π¨. This database covers key data on hardware used to train AI models during the deep learning era. So far, we have collected data on over 100 accelerators (GPUs and TPUs). Read on for some key insights! π§΅
16
18
174
'Is this GPT-4.5'-like questions. ChatGPT doesn't know its model name unless it's in its system instruction or its training data (which is normally many months old), so when you ask ChatGPT something like, "Is this GPT-x?" youβre not exactly encouraging the model to think
18
8
170
MY 2025 AI PREDICTIONS. Agents:. - Agents will finally become useful, releases from all major AI labs, Anthropic will dramatically improve computer-use.- Remote desktops will skyrocket with agent use. - Agents talking to agents will happen.- Personal assistants (customised AI)
52
14
168
GPT-5 / Orion / o3 nearing release?. The "whale" is at the 'staging' point, so could be soon.
The upcoming frontier model from OpenAI appears to be in the staging phase, potentially labeled as O3.
10
10
160
'It cannot be done'. Ha, @MatthewBerman tried to get o1-pro to clone Doom using HTML only: π€£π€£
15
8
156
OpenAI's Sora LEAKED!.
13
12
160
I love the OAI trolling.
7
4
155
@BenjaminDEKR I've never understood the race to the thinnest device. I rather it be thicker with more battery.
5
0
140
o3 is "not that long" away and will come to Plus plan, API prices will be "not that much". I asked about message limits, nothing back a yet.
13
10
145
Star Wars fashion photography. The idea came from an Explore image on the Midjorueny alpha website. Prompts in the ALTs.
7
13
139
o3 is probably the 'merge' of the o line and the GPT line and is the frontier model they started training earlier this year. Wild guess.
22
1
139
If Sam Altman turns up on the next two OpenAI live streams, we'll know something big is coming. π.
17
5
131
You cannot get ChatGPT Operator to use ChatGPT Operator, darn OpenAI, hehe. π
30
8
134
I've not been this excited about AI since first trying ChatGPT nearly two years ago.
19
3
126
@OfficialLoganK No offence here, but it's Perplexity; I use it all day for searches. Second is Midjourney. Then there are about 100 in 3rd, lol.
3
0
124
Thursday 17th is going to be great!. We might or might not get Claude 3.5 Opus, we might or might not get something from OpenAI (GPT-4.5o, SearchGPT, GPT-4o update, o1-full) but what we _definitely_ have is:. - TEDAI Vienna starting, lots of great speakers. - WorldCoin live.
5
16
118
Nvidia's RTX 5090 uses neural nets to generate 90+% of pixels you see. Ray tracing only generates 10%. Games are already being rendered with mostly AI. Crazy.
6
11
120
o1-mini and o1-preview now available to all paid tiers. Streaming is available too.
Streaming is now available for OpenAI o1-preview and o1-mini. π And weβve opened up access to these models for developers on all paid usage tiers.
6
6
117
o1 full will be a big jump, we might even see this next week. But remember, we are still not at the "truly economically disruptive" stage. That is still to come.
o1 will be a big step forward, not just a slightly better o1-preview, and it will have new capabilities from the start. When I said it was a downer that they were considering calling o1 GPT-5, I was mistakenly thinking about o1-preview, but the real o1 actually feels like a big.
13
5
113
Hi @sama, let's #freeo1 today!. @iruletheworldmo has got to get at least one of his dates right. π. And @apples_jimmy wants his Christmas gift early, the patience cave is getting mighty cold. π.
19
9
112
Strawberry out this week, GPT-4.5 out in October, GPT-5 out in December/early 2025. As Jimmy says, put Strawberry into the proper perspective. It is just a stepping stone to the bigger (much bigger!) picture. Don't expect huge things from Strawberry; look at it as a teaser for.
Ok back to October now. We should have a 4.x model ( maybe still called 4.5, my old friend ) in October. The big boy gpt 5, Iβve heard as early as December but for your sanity I would have Q1/Q2 2025 just in case. View Strawberry in context of better models to come.
10
15
111
Lady Vivienne meets Dave the Geezer: Chat #1 -computer problems. My first attempt as an Advanced Voice Mode ChatGPT duo-chat. Sorry, the video quality isn't great!. Enjoy. :-)
15
19
107
It seems to be using Gemini Flash? Here is me continuing the chat, and I've screenshotted the previous response to show it just flipped:
3
3
109
The AGI Countdown went up to 84%. Alan increased it by 1% after DeepMind released Genie 2.
7
11
108
Great, o1-pro is finally working for me. They just took my Β£200 for the new cycle. A day early as well so I didn't make the most of the extra 10,000 Sora credits. Hmmm
17
2
107
The very long awaited Advanced Voice Mode with Vision is coming within the 12 days of OpenAI (next 2 weeks). This was demo'ed ages ago.
Advanced Voice Mode with Vision will be releasing for the 12 days. This is my most sought after feature!.
21
1
106
Gemini 2.0 is just a "few weeks away from the wider rollout":
@SullyOmarr We are a few weeks away from the wider rollout : ).
4
5
107
BREAKING π¨OpenAI Collaborates on First In-House AI Chip with Broadcom and TSMC. OpenAI is developing its own AI chip in partnership with Broadcom and TSMC, aiming to boost efficiency and meet rising infrastructure demands. This project, which scales back its initial ambitions
10
12
106
@basedjensen @harambe_musk So, it's a technique to stop the LLM when it finds varying answers (which suggests it's unsure) before giving an answer.
2
2
103
Who says AI is slowing down? Prrrf. 2024 could still be a hectic year!. Image from The Memo via @dralandthompson
2
21
104
AGI Countdown sits are 81%!. After OpenAI's o1 release, Alan pushed the needle up a whopping 5%! What could the full o1 push it up to?.
Countdown to AGI jumps dramatically on the OpenAI-o1 models . Alan's (@dralandthompson) countdown has jumped from 76% in July/August to 81%! . The 5% increase seems to be the most significant single jump since Sora was announced in February, seven months ago. AGI is coming!
11
5
100
OpenAI have a "November surprise". π.
@jachiam0 Just wait for the βNovember surpriseβ.
12
6
98
I've had a little play with @yoheinakajima's BabyAGI and, with Internet access, feel this could be the best auto-AI. I gave BabyAGI an open objective of producing a 'worldwide accepting AI principle list' and it was fascinating to watch it create task lists and complete actions.
5
3
98
What just happened? πππ. Just summarising tonight's non-Google event event:. After over a week's worth of hype from the @iruletheworldmo account, it resulted in an X Space:. Where @mckaywrigley admitted involvement (please correct me if I'm wrong again,
So @iruletheworldmo has started a new Spaces:.
12
11
92
A new era started today. Gemini 2.0 Flash Experimental was not what I had in mind for the beginning of the new frontier era, but here we are. π. Welcome to the next LLM generation, everyone.
6
2
91
Advanced Voice Mode **Live Video** Rollout Soon?. The following code has reportedly been found in the latest mobile version of the ChatGPT app (v1.2024.317 beta build). It indicates something called "live camera", which could be the live video feature we saw earlier in the year.
ChatGPTβs Live Video feature could be inching towards broader rollout
11
9
93
OpenAI just dropped Canvas in ChatGPT. Think Artifacts from Claude and a bit of Pages from Perplexity rolled into one. Rolling out from today to Plus and Team users, other plans to follow. Check this demo:
Weβre rolling out an early version of canvasβa new way to work with ChatGPT on writing & coding projects that go beyond simple chat. Starting today, Plus & Team users can try it by selecting βGPT-4o with canvasβ in the model picker.
7
7
90
Is it just me is ChatGPT o1 *really* fast rn?. After noticing its speed from a work question, I ran a test below on both o1 and o1-mini, and o1 is lightning fast compared to o1-mini. Sorry, just a screenshot below, not a video, can create one after work if it's still fast.
24
4
90
OpenAI is working on a project called "Perhonen". The word is Finnish and means butterfly. It could be related to metamorphic AI. See comments below for more info.
OpenAI has been actively developing Project "Perhonen" for an extended period.
15
7
88
ChatGPT Canvas Observations + Wish-List. Firstly, I love it. It is easily the biggest update to the ChatGPT UI since its launch. I can sell this to non-users way more easily than before. Previously, non-ChatGPT/AI-users were put off by the 'prompting'. Tools like this will bring.
OpenAI just dropped Canvas in ChatGPT. Think Artifacts from Claude and a bit of Pages from Perplexity rolled into one. Rolling out from today to Plus and Team users, other plans to follow. Check this demo:
7
9
87
I've never used Search GPT but @perplexity_ai is an excellent AI search engine and has replaced Google for 95% of my searches. You can ask it a question and include the style of response, even given it very long prompts. It's my default for my desktop browsers and if I could
Anyone who couldn't get access to OpenAI SearchGPT just like me. Don't be sad, .Perplexity is all you need π¬
11
4
85
Upcoming AI Model Releases Database. @legit_rumors and I have created a database of upcoming AI model releases, including both anonymous models from lmarena and official ones. It includes release ETA, lab, source, and reference links. The link is in the first comment below,
9
16
87
Microsoft has prototypes of near-infinite memory. Expecting it to come online in 2025! π€―. Note to all: make sure to be nice to your AIs now, because they are going to remember everything very soon. π.
We did it. We finally did it. Infinite memory. Microsoft AI CEO Mustafa Suleyman: . βWe have prototypes that have near-infinite memory. And so it just doesnβt forget, which is truly transformative. (. ) 2025 is going to be truly.transformativeβ. Eric Schmidt (Ex-Google CEO)
8
4
85
Sam Altman to present advancements in AI at a closed-door briefing for the US government on 30th January. It's believed the advancements will be "significant".
π¨π¨Scoop: We've learned OpenAI CEO Sam Altman β has scheduled a closed-door briefing for U.S. government officials on Jan. 30 - with people inside and out of govβt telling us AI insiders believe a big breakthrough on PHD level SuperAgents is coming.
11
8
85
@EHuanglu Why "AGI" Mocap though?. Are we going to get a flood or products using the "AGI" term incorrectly not? Or does AGI stand for something else?.
4
0
85
OpenAI is shipping like crazy. Voice is coming to the web, at last!!.
7
6
83
Some interesting thoughts:.
@koltregaskes 3 possibilities i can think of :. a) something that was loaded into the chat history but is not present in the shared conversation skewing what we see and that unseen file or message at the time triggered the response. b) some way to edit the message perhaps a small hack, then.
4
1
81
I think we can safely say the "GPT-x" names are dead. What we'll get is:. o1 [previously GPT-4.5 October?].o2 [previously GPT-5, probably January?].o3 [rumoured GPT-6, 2026/7]. Let's hope they don't do:. o1.o1.5.o1.5o. .
Project Orion. By now we know that @sama loves to spread cryptic messages. And to be honest, I even like solving the riddles a bit. @flowersslop has a good first thought on what Sama's winter constellation might allude to: Project Orion. This is presumably the next big model,
24
7
80
GEMINI 2.0 LAUNCH THREAD:. Logan officially announces Google's Gemini 2.0 model, which will be widely rolled out next year. Gemini 2.0 Flash Experimental:. - 2x faster than 1.5 Pro.- Native image generation.- Multilingual native audio output.- Native tool use.- Multimodal Live
Announcing Gemini 2.0 Flash, a new multimodal agentic model (initially experimental), with our best-ever results on capabilities and benchmarks: .
5
4
81
No GPT-5 today. I might be eating my words in a few hours but the next frontier model seems a few months away yet. It's looking more likely to be a Voice Mode alpha launch, if anything.
12
3
77
Everyone, ChatGPT Tasks are NOT Agents. Separate things. at least for now.
15
2
83
Figure 02 robot is coming!. Now imagine this with ChatGPT's Advanced Voice Mode. ;-).
3
8
80
o3-mini and o3-mini-high are coming soon on ChatGPT from OpenAI:
BREAKING π¨: OpenAI is about to announce 2 new reasoning models soon: "o3-mini" and "o3-min-high". "o3-mini-hype" πππ
5
4
81
All 3 compute levels to be released for 03-mini, so:. o3-mini-low.o3-mini-med.o3-mini-high. More crazy model names, everyone!! π€£
@kimmonismus and we'll have o3-mini high (all three levels will be released same time).
5
4
80