e/acc meat proxy | Physics → ML | Apparently AI researcher nowadays. shitposts, cognition and consciousness
- Tweets 24,231
- Following 1,174
- Followers 1,490
- Likes 35,049
Pinned Tweet
Interesting insight from a niche problem 🧠 ♟️
I finally made my ReCoN learn from scratch! Struggled for a long time where the network seemed *almost* always find checkmate in King rook king, but missed 2/32 positions in about 30% of my runs. Turned out to be in the corners of the board every time.
This led me into the path of trying to understand "what does the network see?" compared to us. Turned out quite interesting and also shed some light on internal representation/"world models". I'm not claiming subjective experience of course, but we do have a representation of how this ReCoN network "sees" its "world" around itself, and how it can benefit (or shoot itself in its proverbial foot) from that.
(Later this led me into a long investigation on representation, boolean expressions and retention of knowledge while staying plastic. but that is a story for another day).
1
6
304
Chips, Coca Cola, chess, ChatGPT and Conciousness discussions.
All these Cs make for a great Saturday night
4
293
MancerAI retweeted
As we all know, free speech is everyone's constitutional right, but it also causes a lot of harm. But that does not have to be a problem. Here is how to solve all issues with free speech! Why not deal with the 1st amendment in the same way as we deal with the 2nd in California?
So, absolutely, you have the right to free speech, as long as you:
- Are at least 18 yo, not a felon and there are no previous speech complaints about you, especially from spouses and neighbours
- Bring 3 good character witness reports, at least one from your closest relatives
- Get through the background check
- Fill in an application, which is different in every county, takes a few months to process, and costs about 1000 USD
- Take an 18 hour course on speech safety (renewable every two years with a 6h class)
Now you can exercise your constitutional right to free speech, with a few reasonable restrictions:
- The speech cannot be anonymous and must always be traceable to you
- No speech in public buildings (schools, libraries, government buildings, hospitals, universities), parks and adjacent parking or business establishments where the owner does not want you to speak
- No speech can be longer than 10 words at a time
- For things that you want to say in public, you cannot use amplification or repetition devices
- For things you want to say in private, the state has an acceptable list of things you can say; the list has not been updated in years but when new things to say are invented, they can be put on the list in the future if deemed safe
- No speech may cross state boundaries; every state has different rules
- You can now go to a licensed speech dealer, apply for the speech you want to make, do the paperwork and wait for a cool-down period of ten days
Someone might have the idea that some things in this list are too restrictive and perhaps not fully aligned with the constitution. That is of course not a problem: you can sue your state and within a few years, the suit may reach the supreme court. If the courts decide to strike down some of these rules, we just make new ones.
48
55
347
15,431
"Matrix multiplication isn't conscious".
Sure, probably not.
Neither is biochemistry.
133
80
1,055
35,791
One former "AI doubter" told me last night "you were right". That was nice. It also made me realize how many other never said that. People that made much stronger claims against AIs future capabilities.
5
116
I am whelmed.
I guess google is (kind of) back in the discussion but I am not expecting any minds to be blown.
Google’s new Gemini 4 Argon equals GPT-6 Astra on the Artificial Analysis Intelligence Index at 60% of the Cost per Task with discounted prices. Google is now back to being one of the top three labs in intelligence achieved
Gemini 4 Argon is @GoogleDeepMind’s first proprietary model above the Flash class in over 7 months. With high reasoning (the highest available), it scores 53 on the Artificial Analysis Intelligence Index, matching GPT-6 Astra (max, 53) and 1 point ahead of GPT-6.1 Sol (max, 52), with gains driven by lower hallucinations and stronger agentic capabilities.
At its current 50% pricing discount and with cache discounts increased to 95%, Gemini 4 Argon costs $1.99 per Intelligence Index task, 60% of GPT-6 Astra (max), but 2.7x GPT-6.1 Sol (max). After the discount ends, this will rise to $3.98 (~1.2x GPT-6 Astra (max)).
Gemini 4 Argon is currently being rolled out to selected users and is not publicly available. The 50% discount is an initial promotion. Google has not yet confirmed the promotion end date
Key benchmarking results for Gemini 4 Argon with high reasoning:
➤ Google returns as one of the top three labs on intelligence: Gemini 4 Argon (high) scores 53 on the Artificial Analysis Intelligence Index, matching GPT-6 Astra (max, 53) and 1 point ahead of GPT-6.1 Sol (max, 52). This is 23 points above Google’s previous non-Flash model, Gemini 3.1 Pro Preview (30) and 12 points ahead of Gemini 3.8 Flash (high)
➤ Launch discounts of 50% make Gemini 4 Argon competitive on Cost per Task: At current discounted pricing, Gemini 4 Argon (high) costs $1.99 per Intelligence Index task, 60% of GPT-6 Astra (max, $3.26) for a comparable level of intelligence. This cost efficiency is driven by lower token prices, rather than reduced token use, with Gemini 4 Argon averaging 62k output tokens per task, compared with 27k for GPT-6 Astra (max). Google has not yet confirmed the promotion end date, but on standard pricing, Cost per Task will increase to $3.98
➤ Stronger agentic performance: Historically a weaker area for Gemini models, Gemini 4 Argon shows improvements across agentic evaluations. It ranks #1 on AutomationBench-AA at 77.5%, 6 points ahead of Claude Sonnet 5.5 (max, 71.3%). On Terminal Bench 4, Gemini 4 Argon achieves 57%, a +53 point improvement from Gemini 3.1 Pro Preview, only behind Claude Sonnet 5.5 (max, 64%), Claude Opus 5.5 (max, 60%) and GPT-6 Astra (59%). On AA-Briefcase, it reaches 1494 Elo. This is driven by a 65% rubric pass rate, the highest we have recorded, but lower Analytical Quality (1576 Elo) and Presentation Quality (1308 Elo)
➤ Lowest hallucination rate among leading models: On AA-Omniscience, Gemini 4 Argon has a 15% hallucination rate, the lowest of any model scoring 45+ on the Intelligence Index, compared with 51% for GPT-6 Astra (max) and 54% for GPT-6.1 Sol (max). This means Argon is much more likely to acknowledge when it does not know an answer rather than guess incorrectly. On accuracy, Gemini 4 Argon scores 50%, a 5 point decrease from Gemini 3.1 Pro Preview, and 13 points below GPT-6 Astra (max, 63%). With this slightly lower accuracy, its overall AA-Omniscience score of 42 remains in line with GPT-6 Astra (43) and GPT-6.1 Sol (42)
Key model details:
➤ Context Window: 1M tokens
➤ Multimodality: Text, image, video, and speech input, with text output
➤ Pricing: $4/$20 per 1M input/output tokens at standard pricing, currently discounted 50% to $2/$10. Cached input tokens receive a 95% discount ($0.10 per 1M at discounted pricing), up from 90% on Gemini 3.8 Flash
➤ Long Decode Continuation: We tested Gemini 4 Argon with Long Decode Continuation, a new Gemini API feature that pauses long responses and resumes them across follow-up calls. This lets reasoning run up to 1M output tokens without request timeouts
1
2
140
The most famous professional flatulist in history, Joseph Pujol (known as Le Pétomane), was a massive celebrity at the Moulin Rouge in the 1890s - right during the dawn of recorded sound
2
82
This was the very instant wherein Zuckerberg beheld his fortune writ large, and knew himself the victor in the race of artificial wit.
For OpenAI and Anthropic, sore pressed, did cry aloud for statutes and edicts, tokens of speech being made common as base coin by the open and unbound alternatives. Their only recourse was to cheapen ever the price of such tokens, that they might still contend. Yet herein lay the bane: they had already pledged hundreds of billions in treasure for engines and forges yet unbuilt. At a certain rate, the return upon that capital doth vanish into air, and the reckoning yieldeth naught but loss.
Thus came they to this cold discovery: that their trillion-fold estimations might be sustained only by one of twain courses. First, to entrench themselves by law, barring all rival models from gaining share. Second, to climb the market’s height and weave themselves vertical, building platforms in the layer of applications.
In the second path hath Zuckerberg the greater vantage o’er both those frontier houses. He holdeth already a mighty host of users...nigh a third part of all the world as monthly frequenters. He possesseth sundry platforms that are fountains of gold, and ask but little fresh outlay each year. Were Meta to cease all spending upon its family of apps tomorrow, those same platforms would yet pour forth billions in free coin year by year.
In essence, he may underwrite the losses of any platform of wit, such as Muse, longer than any rival can endure. This granteth him leisure to bleed Anthropic and OpenAI dry, most of all in the consumer-facing layer of applications.
That Trump would not yield to the petitions of Dario and of Sam is a mighty conquest for Meta and for Zuckerberg. And that is why he cannot stay his smile.
2
103
HOLY SHIT 💥
(That's how you start posts about new AI model benchmarks right?)
Ok Gemini 4 Argon looks really really good on paper. Need to test. Gemini has looked amazing on paper many a times before (granted, it's been a long time since though) but user experience tends to disappoint. Nevertheless it's very nice with many competitive SOTA models competing as opposed to a one horse race.
2
224