Pinned Tweet
my personal litmus test is asking ML engineers whether this is hardware or software
597
76
3,048
269,196
there are probably a few hundred people who have more money than this in assets that are not publicly reported
15
64
14,002
464,532
I don’t get this take, what is wrong with using Android? what OS where you expecting ?
175
69
5,527
875,104
Do you remember which country recently announced no copyright enforcement for AI training anon? 🇯🇵
40
166
2,812
315,782
as a man I have to say this post feels exclusionary, look at this gorgeous hardware I want all of it 😂
10
24
1,795
103,531
he also has the best blog on the web mvdirona.com/
24
65
1,937
567,844
it's WALDO3.0 release day!
the SOTA in FOSS AI for drones gets an upgrade 😁
grab the models here:
huggingface.co/StephanST/WAL…
support further development here:
6228189440665.gumroad.com/l/…
15
124
651
49,734
You need to understand that people want industry that works, and regulation can be required to make sure this exists… but nobody is clamouring for regulation *in the absence of industry*.
4
7
572
48,205
Update: after a few hours of weekend-hacking I've got
@OpenAI's chatGPT using all of English-language
@Wikipedia by providing the text of all articles as source material using a huuge @pinecone
vector database!
I'm happy to report that it successfully answers @nntaleb's question about the battle of Amioun without any bullshitting, as my previous local test did... (this one nitter.app/StephanSturges/s…… ) and it also works in a much more graceful manner as a plugin because the "agent" which you are conversing with can make the decision to source / not source additional content based on context.
The structure that @openAI has built for plugins is really clever and opens up a world of possibilities...
I really think the future of LLMs is in chaining interactions with external sources, and vector DB ones like this that are controlled, maintained and updated by a trusted source play a huge role in making that a valuable system. Stay put for a link to the demo plugin ASAP 😄
@bentossell
And don't forget to support the Wikimedia foundation for all the great content donate.wikimedia.org/
So this is fixed now, @nntaleb: I've added ALL of wikipedia to chatGPT using a @pinecone vector database ! 😄
The AI may not be "smart", but now it has about 58,049,701 pages of knowledge to draw answers from.
... public link coming very soon.
17
59
540
285,779
if you want a pic to look like this it’s mostly about having enough flash to over power the sun by about 2 or 3 f-stops
2
5
449
400,664
And now... your drone can see 😄
Perhaps the last teaser for WALDO 3.0 before release.
Version 3.0 is built on a YoloV8 backbone and features inference examples using the fantastic Supervision annotation and inference system from Roboflow (thanks @skalskip92 )!
More coming very soon.
6
39
456
43,212
Releasing NANO: a very small and lightweight detection network for drones that detects a few very simple classes in low-rez.
github.com/stephansturges/NA…
6
87
433
45,613
I think it’s obvious: the @dieworkwear x @ahm_therebel’s granny capsule collection coming Fall 2025.
1
339
20,424
I'm finally getting around to releasing the Unidrone FOSS AI models for drones, check them out here:
huggingface.co/StephanST/uni…
Sorry for the delayed FOSS release, it's hard to find any free time these days to work on these kinds of things... anyway let me know what you think 😄
12
61
353
27,566
"words of wisdom" and it should be a 30sec add-on at the end of each "clarkson's farm" episode where Gerald is sitting with a beer in front of a lovely sunset imparting the lessons he's learned along the way to the next generation
5
310
5,683
day 3 of dream window: the prompts are now engineered by chatGPT allowing input in any language that is understood by Whisper + gpt3.5 turbo, and the chatgpt bot has been taught to engineer excellent prompts for images 😄
@OpenAI @bentossell @jposhaughnessy @StabilityAI
18
27
295
83,132
So this is fixed now, @nntaleb: I've added ALL of wikipedia to chatGPT using a @pinecone vector database ! 😄
The AI may not be "smart", but now it has about 58,049,701 pages of knowledge to draw answers from.
... public link coming very soon.
We're a loooooong way, AI.
When asked about the battle of Amioun, which plotted Greek-speaking Byzantines ag. Aramaic-speaking Monothelites, GPT-3.5 made up a story of Islamic invasion in 632, based on the general history.
I tried with GPT-4, which said "No such battle".
22
39
300
194,515
This is how much it costs to calculate embeddings for ~230K wikipedia articles using @OpenAI's API.
Incredibly cheap.
16
20
309
66,237
Demo: one minute of UAV visual navigation from satellite imagery where GPS coordinates are derived by comparing an image from a downward facing camera to a prior satellite image 😄
This is a solution for high-altitude navigation with no GPS.
25
37
356
74,152
the alternative take is: “I want a strategic role because I’m technical but i’ve seen too many good projects fail because the strategy planners were not technically competent. Give me the strategic role and this time shit will work.”
8
6
229
Sunday funday project: I'm making wikiGPT: it's chatGPT with all of english-language Wikipedia injected as context using a @pinecone vector database.
People keep asking me what the point of context injection is. Here are two pics that explain it really simply 😄 1/n
17
22
205
36,744
WALDO v2, video 2!!!! 👀
Not 100% perfect... but anyone can pull the code and draw their own conclusions so there's no incentive for bullshit! 😁
sign up on aircortex.com/ to get early access to the model and code
23
28
189
35,393
I feel this is less about wokeness than about the trillion-dollar non-profit sector making sure you know they are still needed and won’t be replaced. Although that’s probably also where a lot of the woke rhetoric is created in the first place.
10
5
188
37,670
they didn’t let him in because he also invented the quad damage power up and people got nervous
1
168
17,794
Video preview of the W.A.L.D.O. v2 beta 👀
While it's still not perfect on all these test videos, it is a huge step up from W.A.L.D.O. v1 which was released just 2 months ago!
Coming very soon to your UAV / EO workflow!!!
Sign up for access at aircortex.com/
7
15
146
19,054
There is a tremendous arbitrage opportunity in setting up a datacenter in japan and scraping all the internet's data which is under copyright elsewhere.
Raising 100M to get started, who is in?
Japan recently reaffirmed that it will not enforce copyrights on data used in AI training.
The policy allows AI to use any data “regardless of whether it is for non-profit or commercial purposes, whether it is an act other than reproduction, or whether it is content obtained from illegal sites or otherwise.”
HN Discussion: news.ycombinator.com/item?id…
Article: technomancers.ai/japan-goes-…
(h/t @Dan_Jeffries1)
ALT I strongly agree with this. There's a distinction between "learning from" and "copying". "Learning from" is a transformative process that distills from the observation. This distillation can be as simple as indexing for a search engine, or as complex as a deep neural network. Simply because a neural network can create something that is a copyright violation doesn't mean the training process itself it. A human can see a advertisement for a Marvel movie and then reproduce the Marvel logo. Redistributing (and possibly actually doing that reproduction) that logo is a copyright violation, but the learning process isn't. The neural network is a tool. It's reasonable to be concerned about the loss in employment by people who are affected by generative AI. But I think this is a separate issue to the copyright argument.
6
5
136
35,956
this implies the existence of a terrible-looking and overpriced house for sale in Prognac.
1
120
20,668
My brother in Christ if you thought that was bad 🤣🤣🤣😄
lemonde.fr/en/europe/article…
1
9
117
16,842
very quick update on this thing 😁 @OpenAI @StabilityAI
8
18
126
15,697
"Hey how many cars are in our parking lot right now?"
Class 'car': 4046 occurrences
Class 'van': 7 occurrences
Class 'truck': 5 occurrences
Class 'building': 2299 occurrences
Class 'human': 122 occurrences
Class 'digger': 1 occurrences
Class 'container': 6 occurrences
Class 'bus': 20 occurrences
Class 'pylon': 61 occurrences
Class 'boat': 30 occurrences
Class 'bike': 173 occurrences
Class 'solarpanels': 12 occurrences
8
10
129
17,470
An update on my satellite image-based visual positioning for high-altitude autonomous navigation!
We are simulating a drone flight at ~1000m altitude in this demo with a 35mm equivalent focal length camera, and the input map is 0.7m per pixel.
4
21
120
12,875
Statistically there might be a guy like this out there who has not yet been laid off from a single one of those jobs 😆
108
24,653
people keep saying this but have you tried actually living the Android phone / galaxy buds / google watch / Quest VR lifestyle 😅😂🤣?
9
110
16,812
HOT is a nano-sized YOLOv8 model that detects people in thermal imagery. Grab is here:
github.com/stephansturges/HO…
6
18
114
12,142