Pandering Panda

@jeremyct The problem is that what you want is DEFLATION, not slowing inflation (what we got.) You want deflation… until you realize the implications for the unemployment rate and your portfolio. There’s no going back.

My wife and I got into it last night about the grocery bill. I pulled up an old Target order from February 2020. 28 items. $64.50. Added every single one back to the cart just to see. $158.30. Same 28 items. Same store. Same cart. $93.80 more. In five years. We didn’t buy more. Didn’t upgrade anything. Didn’t add a single extra item. Just needed the same things we always needed. They told us inflation was transitory. They told us it was under control. They told us the economy was strong. Our cart didn’t get any of those memos. 145% increase in five years. Not one member of Congress has looked at a grocery receipt in decades. And it shows.

Are you all working on re-running these with fable, GPT 5.6, and grok 4.5? Anecdotally, I find fable's writing/research is stronger than opus, but it still often makes contradictory claims that it doesn't disentangle/grapple with. GPT 5.6 Sol pro often goes much deeper and creates more coherent research reports. This is a small sample size (n ~ 10), but every time I've used an identical prompt between fable and 5.6 pro and then fed the response from each model to the other, fable admits that it missed important facts and alters at least some of its conclusions based on the evidence from 5.6. This may also be a harness or search capability issue.

I think you're right -- they have made some progress. For example, on the NewsBench neutrality measure from Forum AI (where I'm an advisor) they do show a right-wing lean while the other models lean left. (#neutrality" target="_blank" rel="nofollow noopener">byforum.com/newsbench#neut…) I also believe they've optimized more for decisiveness than right-lean, and at least last I saw, Grok scores much higher in terms of lower rates of refusals. So I was wrong to suggest they haven't made progress---but on a number of different tests of slant, like the WaPo one, they continue to score as left-slanted, which suggests there's still significant room for improvement. At the same time, Grok's goal isn't probably to score like Gemini does on neutrality. Gemini nails that score by relentlessly hemming and hawing, while Grok wants to be decisive.

The models lean left in large part because they are trained on huge amounts of text from the open internet that was disproportionately created by people with left-wing views. You can't reverse this lean through ex post prompting alone: it requires effort and money. Gemini is the most neutral probably because Google poured immense amounts of resources into specialized training to "undo" this lean. SpaceXAI is a much leaner company, and hasn't put those same level of resources into changing the lean of the model despite their public commitment to that cause.

@blondesnmoney Did you have to delete the OP because you were leaking too much alpha?

@CL207 What’s your ELO?

just got a city on geoguessr that looks like some kinda identity crisis

@copiumfueled You realize the admin could have, checks notes, ignored Dario and just let it rip? Or listened to ANT when they said Fable is safe enough for general release? Blaming it on Dario is just convenient because everyone involved knows they must not criticize the dear leader.

The two best AI models in America are sitting in a drawer right now. David Sacks explained why this week on All-In, from inside the administration. China just dropped GLM 5.2. Open weight, MIT license, free to download anywhere. It beats GPT 5.5 on coding and trails Claude Opus 4.8 by under a point. The best openweight model on earth, and it came out of Beijing. Fable got rolled back. GPT 5.6 is stuck navigating new approval hoops. Both frozen by a jailbreak report and a safety fight nobody outside Washington fully understands yet. So the most open country ended up with the closed models, and right now they are off the market. Sacks traced how America got here. Dario spent months lobbying for a federal AI regulator. He wrote a blog post asking for an FAA for AI. A government approval process for model releases. And he got it. The administration moved, and the first thing the new caution did was freeze his own company's model. The Chinese model has no such problem. No approval queue. No regulator. Trained on Huawei chips, packaged as "AI in a box," sold globally at a fraction of the cost. Sacks has been saying it for months. We invented reasons not to ship. We defined this as a race and then handed the other side a head start. What everyone is pointing at is real, but none of this was fate. A freer society didn't take the open models. America shut its own best models in a room and called it safety. The closed models come from the open nation because the open nation chose to close them. Watch the whole thing on @theallinpod

@EgeErdil2 The OP mentioned “foreigners,” but the only country that really matters is China because they have the industrial scale to utilize that knowledge. You only spy for China because they are paying you or you have national allegiance. That was not the case for the USSR.

@EgeErdil2 A relevant fact is that the government they were leaking secrets to was broadly considered the embodiment of far left-wing beliefs/communism. At the time they leaked, the USSR was a key US ally (Uncle Joe).

holding far left-wing views was a much stronger predictor of spying risk during the manhattan project than not being a US citizen every known spy was a communist or had strong left-wing sympathies, and almost all were US-born

@HedgeDirty so how do they feel knowing that a bunch of OAI/Anthropic empolyees in SF who wear company t-shirts and rent $4k/month 1bd apartments in Haight-Ashbury because they can't be bothered to move have a higher net worth than them?

People in finance do not dress as they do because it's "part of the forgiveness for wealth in finance" People who go into finance do not believe wealth is something that requires forgiveness at all They dress as they do because *absolutely everything* is a competition to them

@buccocapital This is exactly why it was so easy to long $GOOG at $80 when BG2 and every tech podcast was talking about using LLMs for everything and never using Google. Helped explain just how out of touch tech investors were with how normal folks operate.

I genuinely don’t understand this: 1. Google AI Mode has access to tons and tons and tons and tons and tons and TONS of sites that ChstGPT and Claude can’t access 2. If you’re in Chrome it is one million times easier to use the search bar for a query 3. It is objectively better for all local and navigation searches Not using Google anymore is literally just using the internet in a worse way. Not even picking on Signull I’ve seen plenty of “tech forward” people say this. It makes no sense to me.

@dampedspring to be fair, binance and other "international" crypto exchanges have been operating bucket shops for years now and are some of the only successful businesses within crypto. BNB is still worth $80B and has massively outperformed crypto beta on almost any time frame you choose.

Should we buy the bucket shop? Makes sense to until all the clients are liquidated

Bucket shops 101 "In the first place, you didn't buy or sell stock. You didn't buy 100 shares of Erie to hold or to sell at a profit... You made a wager on the price, just as you might have done on the outcome of a football game." "You could buy on a margin of one per cent. That meant that if you had twenty dollars you could buy one hundred shares of a stock selling at twenty... If the price dropped to nineteen and seven-eighths you were wiped out automatically." "The margin was so small that the slightest wiggle against you wiped you out. The shop just took your money. It was a mathematical certainty that you'd lose." "The bucket shops... would see that there was a large (long ) interest in some stock. They would combine and (sell) a few thousand shares of that stock on the Real Exchange... and run the price (down) a point or two. That was enough to wipe out all the margins of the customers who were (long)" These quotes are from 1923. Perps are not innovative.

this is why reminiscences of a stock operator is still required reading. that book made the crypto cycle a lot easier to understand. in crypto, "market makers" get given calls by the token originators in exchange for "stabilizing price action" (aka make number go up) so the founders can burn shorts and then exit to some retail or vc bagholders. but what happens to token prices when enough people get burned and those calls are worthless?

@mezmotweets @niccruzpatane It’s just the law of large numbers. It could be money-making to buy and hold SpaceX stock. But a 100x would require it to be worth 2.4x the value of the current total US stock market valuation. Comparing this IPO to the Mag7 IPOs is just non-sensical.

@Pandering_Panda @niccruzpatane you’re so right. the people who are about to buy in are exit liquidity to get the rich richer. next couple of years are going to be phenomenal 😭

Just gonna leave this here. Long-term thinking always wins.

@mattyp @atmoio yeah, this is sama trying to position OAI as the long-term winner over ANT. gpt 5.5 models are so much more token/cost efficient than Opus. if budget starts to matter, then you'll have a lot of enterprises switch to codex/gpt models to save.

@atmoio idk like he's taking a pretty level-headed approach relative to others in the space i wouldn't really bet against this guy

Sam Altman is starting to panic

@gregoryblotnick Yes, he’s really paid the cost for this… Trump still lets him ride along on trips, needs his $ for campaigns, and TSLA is up 80% post election despite fundamentals tanking. Being able to call someone a pedo and then get invited to their house later is a flex.

@sama Let’s go!!!

GPT-5.5 is going to have a party for itself. it chose 5/5 at 5:55 pm for the date and time. if you'd like to come, let us know here: luma.com/5.5 codex will help the team pick people from the replies. 5.5 had some good ideas/requests for the party, which we'll do.

I remember reading “goblins” from GPT5.4 for the first time and then checking my custom instructions to make sure someone wasn’t goofing with me and telling it to add goblins or gremlins in absurd contexts in an otherwise coherent response. I couldn’t take the rest of the response seriously and just went with Claude’s answer.

everyone is assuming this is some kind of quirk chungus marketing campaign but if you’ve worked with 5.4 and beyond they tend to call everything goblins, gremlins etc and it’s just super noticeable and if you work with them all day you start to get annoyed

@Helios_Rises @puffinancial @edzitron Where do the GPUs live is CAPEX (not in the calculation). Everything else you mentioned is OPEX.

@Pandering_Panda @puffinancial @edzitron Where do those gpus live? How much does it cost to house them? How much does it cost to keep them in a safe operating environment? How much is insurance? There are a lot of upstream costs that aren't factored in here.

I’m sorry what? it appears that hyperscalers and neoclouds were taking less than $4 per gpu for hour for B200s - that is insanely low, I don’t even think it’s margin positive. Everyone’s framing this as “massive demand” but said demand is at very low prices

All makes sense but opex seems low if we’re averaging across a whole DC. And you’re not considering the cost of capital and original capital investment to build the DC (electrical, plumbing, permits/legal). If you’re just plugging in new chips to an already constructed DC, that’s a bit different, but still not zero capital cost to retrofit to work with Blackwell.

@edzitron GB200 NVL72 is $3M give or take $42k per GPU closer to $60k all in 4 year useful life is $1.71 at 100% util +10c for power & +25-40c for opex is $2.20