andrksl

swe. thoughts on productivity, llms, markets

[andrksl @andrksl]

@andrewmccalip I strongly suspect 5.5 in the Codex MacOS app is routing to 5.6.

The frontend / UI taste improved dramatically from a week ago.

[andrksl @andrksl]

Oh they are definitely routing 5.5 queries in the Codex MacOS app to a new model.

Big improvements to UI/design-sense compared to 5.5. Seems like 5.6 will finally be on par with (or maybe even better than...) Opus for frontend.

[andrksl @andrksl]

@vladtenev Dumbest move ever was for Robinhood to lean so hard into sports betting and prediction markets before it even provided access to more than a few international stock listings. Cementing itself as a gambling app instead of an investment app.

[andrksl @andrksl]

@powerbottomdad1 AMZN META MSFT are already there or close. SPCX GOOG need a bit more time.

[Steve Hou @stevehou]

I’m bullish but I’m not complacent. It’s def much harder to remain a bull in June 2026 than any of the past 3 years. I think we’re overdue pullback but the catalyst isn’t going to be macro but much likelier AI related such as one of the mega tech like meta pulling back on AI capex or if open AI models resurfacing the question of ROI. I consider AI related risks much more impactful than traditional macro risks which if anything could help lengthen the AI runway

[andrksl @andrksl]

@andrewmccalip When will the option to opt-in to more data sharing be ready?

It was neat at the beginning, but now making just 1 cent an hour kind of kills it. Do you think the long term equilibrium payout has a chance of hitting ~$3+/hour? Otherwise not worth it for most ppl in the US.

[andrksl @andrksl]

If you believe in ASI then you also believe Nvidia’s moat is fragile.

The value should be captured by TSMC and INTC, to fabs not chip design + software.

The frontier lab that develops ASI no longer has a dependency on NVDA.

[andrksl @andrksl]

@AndrewCurran_ This definitely fucks the IPO plans, no?

[andrksl @andrksl]

@bourscheid Other stupidity aside, the government’s annual spend exceeds $7 trillion.

If it were possible to solve all the world’s problems with one extra trillion we would have done that by now.

[andrksl @andrksl]

@andrewmccalip The ads aren't sync'd well. I'd prefer the experience to be consistent. I am seeing three separate ads at the same time, and I can only actually click / convert on one (the Co-Invest).

[andrksl @andrksl]

You don't need superintelligence to make dashboards in React or to draft emails to coworkers.

Once an intelligence threshold is passed, the marginal benefit of using the smartest model is largely gone. Enterprise cost discipline is coming to token-spend before EOY.

[andrksl @andrksl]

Decided to give @antigravity a try. It did not even complete the first request before it told me my quota had been reached. $20/month, "AI Pro" plan lol.

@googleaidevs @OfficialLoganK Do I need to pay $100/month to get a single request through?

[andrksl @andrksl]

@kennymxu I was under the impression that management at Meta got more messed up than the ICs.

[andrksl @andrksl]

@voodoo_bubu The irony of telling a story about teaching English to Asians when you don’t know the difference between your and you’re.

[andrksl @andrksl]

@AndrewCurran_ Global dividends would probably get backlash from Americans. Can’t imagine if this were up to Americans or their representatives that we would vote to reduce the share Americans get by 95%+?

[andrksl @andrksl]

@thsottiaux ur right just don’t tell us next time

[andrksl @andrksl]

@bayeslord “Verifiable vs non-verifiable as a meaningful distinction will fade.”

How?

[andrksl @andrksl]

@Altimor In one year we will not still be paying $25/million tokens for Sally to draft emails to Bob, or for Bob to generate dumb React dashboards.

[andrksl @andrksl]

@Altimor At that point if OpenAI + Ant are not solidly in the era of generating novel, economically valuable scientific work w/the models I really start to question their valuations. 100 IQ desk job LLMs use cases will likely be covered by distilled Chinese models at 10% cost of today's.