Eventually you’ll be right, after missing out on 50% gains.

I remember a prominent poaster loudly proclaiming he was going all cash after the first few months of the Trump run first term.

He retired a couple of years later. The S&P is up 3x since.

(http://www.autoadmit.com/thread.php?thread_id=5783988&forum_id=2\u0026mark_id=5307129#49331484)

ps im retarded and gay

(http://www.autoadmit.com/thread.php?thread_id=5783988&forum_id=2\u0026mark_id=5307129#49331522)

(http://www.autoadmit.com/thread.php?thread_id=5783988&forum_id=2\u0026mark_id=5307129#49331489)

I think the current gains aren’t sustainable. But that might just mean a decade of slow gains.

I am diversifying. Was 99% VOO for the past ten years. But I’m adding some BND and SGOV for diversification. If the market tanks 10% I’ll move back to full VOO.

(http://www.autoadmit.com/thread.php?thread_id=5783988&forum_id=2\u0026mark_id=5307129#49331550)

just keep DCAing into the market. forever

(http://www.autoadmit.com/thread.php?thread_id=5783988&forum_id=2\u0026mark_id=5307129#49331555)

(a) you need the money or

(b) clearly have somewhere better to put it (eg exuent is pushing $8 SOL)

(http://www.autoadmit.com/thread.php?thread_id=5783988&forum_id=2\u0026mark_id=5307129#49332138)

i've always been a big believer in the S&P or total stock market index funds, but i have put like $250k into individual stocks - TSLA, NVDA, PLTR, HOOD, RKLB (thanks poaster). i think these companies are all strong long-term HODLS. i could be wrong, who knows.

(http://www.autoadmit.com/thread.php?thread_id=5783988&forum_id=2\u0026mark_id=5307129#49331769)

energy is likely to only be a major obstacle by around 2030, when training runs are likely to be capped at 10,000x GPT-4 scale.

(http://www.autoadmit.com/thread.php?thread_id=5783988&forum_id=2\u0026mark_id=5307129#49332171)

This becomes even more interesting when applied to multimodal transformations, e.g., video-to-audio (video input generates audio output), video-to-text, and image-to-video. I'd imagine within the decade, we'll see models that can output entire film scores using MoE as a multimodal router: A vision expert for generating motion, spatial, and gesture features, an audio expert for synchronized sound/voice, a language expert for narrative intent, and a style expert to enforce cinematic tone, pacing, and emotion. The gating network would learn to select the correct expert sequence or combination for each transformation stage. Here, RL would ensure that the final perceptual outcome aligns with both factual and aesthetic goals. That's where the industry is partially heading.

Because each modality in a multidomain system has its own geometry (computer vision embeds spatial-temporal tensors, audio embeds time-frequency spectrograms, text embeds discrete semantic tokens, graphs embed relational topologies, etc.), we can't use common reasoning or routing layers to directly operate on all of them. Embeddings unify these spaces into shared manifolds that enable measurement of similarity, causality, and semantic relationships. Among them, graph embeddings are especially powerful for structuring multimodal reasoning to capture contextual causality, thus enabling a CoT reasoner to query or traverse the embedded graph as a semantic memory. If integrated before gating, the gate would see the relational context (i.e., the gate not only sees "visual feature = lightning" but also "this causes sound event "thunder"). Integration during reasoning would allow CoT steps to traverse edges ("because A->B, we expect C") and integration after generation would facilitate validation of cross-modal consistency.

This is particularly powerful because graph embeddings enable compositional reasoning, giving the model the ability to infer unseen relationships ("A is taller than B and B is taller than C, so A must be taller than C"). Moreover, embeddings act as a shared state between experts enabling different experts to read/write to the same latent "memory". When MoE routing becomes embedding-aware, the gating policy can use distance/centrality/clustering in the embedding's space to select experts. Finally, embeddings provide structural support to CoT, because the reasoning steps traverse embedded graphs (symbolic + neural) rather than linear token sequences.

If we define "reasoning" to mean navigating in embedded space via learned expert operators under an RL-optimized policy, then AGI is already here.

(http://www.autoadmit.com/thread.php?thread_id=5783988&forum_id=2\u0026mark_id=5307129#49333034)

The skeptical case is specific to the lack of actual implementation of real, measurable and successful AI deployments in large enterprises. And while we all may grasp that it’s inevitable on a long enough timeline, there’s still as much uncertainty of where we will be in two years as there was a year ago. Cf. public cloud adoption where 1) the business cycle was perfectly timed, and 2) there were these enthusiastic large scale early adopters who had logical and intuitive reasons to take the leap and prefer it as the premise of how they would operate in the future. We are already at a point where companies who were interested in AI implementations tried some evals and have let their enthusiasts spend 6+ quarters hyping their projects while they’ve dicked around. The hope a year ago and currently is that most of those companies are going to follow those evals up with actual spend of dollars in their BU budgets, but many companies are just trying other evals and not putting up any real money.

The current phase of the hype cycle is there are a bunch of bigcorp middle managers who have spent the last 2-3 years super studying AI and hoping that it was their ticket to an exec title and the career success they have craved for the last 7-10 years (since they realized they missed the boat with cloud and SaaS). When their management decides to pull the plug on any real spend—eg because we have 2-3 straight quarters of stagflation and demand for their products is slowing and they want to tighten belts to support the stock price—there will be a correction. I don’t think it will be a huge crash, thus my 10-15% call, basically putting the market back at the trajectory it had from 2010-2019. But I’d call some kind of correction like that more likely than not between now and the midterms (for obvious reasons).

(http://www.autoadmit.com/thread.php?thread_id=5783988&forum_id=2\u0026mark_id=5307129#49332540)

Until a few days ago the sentiment was THIS RALLY IS CRAZSY its gonna crash but then u had a few bigwigs say oh its gonna boom more and now the sentiment seems positive

which prob means its gonna crash but who cares cause jews rule so actually a 10% correction now wld be healthy for QQQ 700

(http://www.autoadmit.com/thread.php?thread_id=5783988&forum_id=2\u0026mark_id=5307129#49332481)

u have to literally time the market perfectly to make money on puts in this market cause shit cld easily just go up 5% more

alt is to keep buying down as u get ass raped and hope it has a down day like today

but just buying a put 3m out? it cld get ass raped if market keeps going way up

(http://www.autoadmit.com/thread.php?thread_id=5783988&forum_id=2\u0026mark_id=5307129#49332554)

That said stocks do have risk, so if you’re carrying more risk than you want, yes of course bring it down. Pay off debts, carry some allocation to cash/bonds, whatever. But there are zero smart rich people who made it by flipping between 0% and 100% equities based on general sentiment.

Edit: I forgot to note that everything is always scammy. Tech bubble, Enron, Worldcom, Lehman, NINJA loans, Madoff, crypto, AI. There’s always something to worry about. That’s why you don’t buy one stock and ride it to zero. But good investments make up for all the scams.

(http://www.autoadmit.com/thread.php?thread_id=5783988&forum_id=2\u0026mark_id=5307129#49333336)