Has anyone tried quantizing some metric of brain activity (EEG etc.) into a set of tokens, and then training a Transformer model on it?
I wonder if hidden in that is some good prior for human-cognition-related activities, i.e. extend the token space to add human language tokens, train on that and see if it trains significantly faster than a randomly initialized model.
I guess no one else is allowed to have indirect lighting anymore (/s)
As someone not well versed in how startups work, why do they need to acquire the entire company to have the team work for them, as opposed to negotiating with the employees individually? Is it essentially a lump sum payment to stop what they were currently working on.
1) have an idea, negotiate seed funding. Someone will give you a bunch of money to make an MVP. the "angel" or "seeder" will exchange money for a percentage of the company. They will have a say in how its run
2) grow the company according to the metrics that make you look valuable (market share is normally the common one)
3) growing costs money, so raise a "series a" same deal as before, either seeder/angels are bought out, or their share of the company is diluted in exchange for more monoey from the funders, typically a VC.
4) repeat until either bankrupt, IPO, Profitable, or boughtout.
Now. As an early employee, you are given shares in the company. Typically <1%. the more rounds you have, the more diluted that percentage gets.
when the company is bought out, your shares are exchanged for a price. however as you didn't put the money in, you are a way down the pecking order for getting the cash. Not only this, but a "buyout" price is rarely the headline price. For example CTRL-Labs[1] was "bought out" for $500 mill. However thats not the price they paid for the company. thats the price paid for the company and the "golden handcuffs" to keep employees.
of that 500mill some of is say <75m will go to physically buy the company, the rest will go towards golden handcuffs to keep the employees working. After all, if you give most of the company >$2m upfront, they'll say "work for facebook? fuck that, I'm retiring"
So to combat that, you'll get your 0.0001% of $75m, then offered a contract with the new company that will give you shares that will be released to you gradually after a certain cool off period, say one year (Fb and google give you 6.25% every 3 months.) That means that to get your $2m you need to work for the new company for four years.
> As someone not well versed in how startups work, why do they need to acquire the entire company to have the team work for them, as opposed to negotiating with the employees individually?
Because otherwise, the team would have the option of continuing their current employment.
Take away the alternative they are demonstrably already choosing over you, and you have a better chance of getting and keeping them.
It also usually triggers a liquidity event, so the team gets paid (or their equity gets transferred to OpenAI) in a way that provides a bit of a bonus and incentive to stick around.
I hear a lot about people talking about politics becoming ingrained in certain scientific fields, but it's hard for me to imagine this happening in certain fields. It is amusing to try to think how one might even begin to "ideologically subvert" a field like Condensed Matter Physics or Knot Theory.
Can you find anyone who would be upset about the Quantum Hall Effect or the Călugăreanu Theorem.
Marxists were upset about relativity and quantum mechanics:
"Soviet physicists were saved from a grand inquisition ironically thanks to the practical power mongering of Marxist dictator. [...] “Stalin chose the bomb over ideological purity.” Josef Stalin had ordered a core group of physicists to create atomic weapons for Soviet Union. These physicists insisted that inputs from the physicists accused of heresy to Marxism, was essential for the atomic bomb project. [...] Beria the killer hand of Stalin asked Kurchatov the physicist whether it was true that quantum mechanics and relativity theory were idealist, in the sense of antimaterialist.
Kurchatov replied that if relativity theory and quantum mechanics were rejected, the bomb would have to be rejected too. [...] three leading physicists -Kurchatov may have been among them- approached Beria in mid-March 1949 and asked him to call off the conference [organized by the party to ensure that physics followed the principles of dialectical materialism] on the grounds that it would harm Soviet physics and interfere with atomic project.
Beria replied to the physicists that he would not make a decision on this himself but would speak to Stalin. Stalin agreed to cancel the conference, saying of the physicists, “Leave them in peace. We can always shoot them later.”
"The privileging of solid over fluid mechanics, and indeed the inability of science to deal with turbulent flow at all, she attributes to the association of fluidity with femininity. Whereas men have sex organs that protrude and become rigid, women have openings that leak menstrual blood and vaginal fluids. Although men, too, flow on occasion — when semen is emitted, for example — this aspect of their sexuality is not emphasized. It is the rigidity of the male organ that counts, not its complicity in fluid flow. These idealizations are reinscribed in mathematics, which conceives of fluids as laminated planes and other modified solid forms. In the same way that women are erased within masculinist the- ories and language, existing only as not-men, so fluids have been erased from science, existing only as not-solids. From this perspective it is no wonder that science has not been able to arrive at a successful model for turbulence. The problem of turbulent flow cannot be solved because the conceptions of fluids (and of women) have been formulated so as necessarily to leave unarticulated remainders."
“To provide an example of the role that white empiricism plays in physics, I discuss the current debate in string theory about postempiricism, motivated in part by a question: why are string theorists calling for an end to empiricism rather than an end to racial hegemony? […] Given that Black women must, according to Einstein's principle of covariance, have an equal claim to objectivity regardless of their simultaneously experiencing intersecting axes of oppression, we can dispense with any suggestion that the low number of Black women in science indicates any lack of validity on their part as observers. […] In string theory, we find an example wherein extremely speculative ideas that require abandoning the empiricist core of the scientific method and which are endorsed by white scientists are taken more seriously than the idea that Black women are competent observers of their own experiences.“
In any case I imagine that trying to create a commerce "everything" app like WeChat in the US will cause trouble with Apple and Google, even Epic Games w/ all their Fortnite users couldn't dodge the 30% fee.
Late to the party, here is some code to solve the path between each word, just requires numpy and a txt file of 4 letter words "words.txt" separated by lines
import numpy as np
with open("words.txt") as f:
words = f.read().splitlines()
f.seek(0)
chars = np.array(list(ord(c) for c in f.read() if c != '\n')).reshape((-1, 4))
word_to_index = dict(zip(words, range(len(words))))
char_flip = np.sum(chars[:, None, :] != chars[None, :, :], axis=2) == 1
chars_sorted = np.sort(chars, axis=1)
char_shuffle = np.all(chars_sorted[:, None, :] == chars_sorted[None, :, :], axis=2)
adj = char_flip | char_shuffle
np.fill_diagonal(adj, False)
# import itertools
def path_recursive(current, end, adj, path, depth):
if depth <= 0:
return []
if current == end:
return [[*path, end]]
# all_paths = []
for neighbor in adj[current]:
if neighbor in path:
continue
_path = path_recursive(neighbor, end, adj, [*path, current], depth - 1)
# all_paths.append(_path)
if _path:
return _path
# return list(itertools.chain(*all_paths))
return []
def binary_shortest_dist(start, end, adj):
depth = 1
adj_n = np.eye(adj.shape[0])
while depth < 20:
depth += 1
adj_n = adj_n @ adj
if adj_n[start, end]:
return depth
def path(start, end, adj):
start = word_to_index[start.upper()]
end = word_to_index[end.upper()]
depth = binary_shortest_dist(start, end, adj)
print(depth)
adj = [[i for i in np.nonzero(_adj)[0]] for _adj in adj]
paths = path_recursive(start, end, adj, [], depth)
return [[words[i] for i in path] for path in paths]
path("disc", "zero", adj)
Or before the images accompanying a comparison article between a Samsung Scene Optimizer moonshot and $4,800 Sony astrophotography rig become completely incomprehensible:
Don't want to bring too many Reddit memes over, but if the browser enhances the images on both sides of the comparison, and corporate wants you to identify the differences... what can you say but "they're the same picture"?
Just to be clear, Samsung was caught using basic pattern matching to detect the camera being pointed at the moon, triggering wholescale substitution with the image from the $4.5k camera. A redditor blurred an image of the moon, took a photo of the blurry image on their screen, and got the high-res image.
That's right. They use a fully differentiable renderer, which allows them to optimize the properties of this set of 3D gaussians using back propagation / gradient descent.
But ultimately what they end up with is an explicit representation of the scene, so unlike NeRF there's no "inference" during rendering. In that sense it's somewhere between a traditional mesh representation and an implicit representation like NeRF or signed distance fields. That's what makes it fast too; they get to make use of the rasterization acceleration capabilities of GPUs , unlike NeRFs which need to be sampled many times along a ray to render a scene.
