gazingback

@TheSpeculator0 probably hard to fix fast cause it’s a compute capacity issue - and they were “reasonable” in their forecast and purchasing

Can anthropic stop shipping 100 new peripheral features per day and just make sure claude code works thanks

@FundamentEdge Anthropic is more bitter lesson pilled and that will matter in near horizon - gpt is currently better at analytical work due to a lot of effort and focus, but it makes trade offs in terms of ease of use and general capabilities in other work settings

Vibes are very pro-Claude right now, but it does seem that OpenAI has all the pieces to build an incredibly good “Cowork”style agentic work platform: > GPT 5.4 Pro (very good at Excel) > The best Deep Research capability > Codex (a better one shot coding agent) Investment Process Deep Research running 24/7 365 on 3 Key Drivers, 2 key Risk Factors on every name in coverage with Excel ingress/egress with exoskeleton research support where needed (often with the human orchestrated by the LLM). Ideas for Investing AI that felt conceptual only 3 months ago feel like they could be very close to an operational reality. It’s a strange time to be alive. What vendor wins? I don’t know, but the prior thus far has been leapfrogging models, fast cycle feature parity (I.e laggard copies the winner), and the continuous abstraction of harness engineering, which I think should be considered carefully before making large scale institutional decisions that are annoying/costly to reverse.

@bcherny Quantum physics professor experience

A weird part of working at Anthropic: getting a few of these each day

@tanayj @_sholtodouglas @mattturck @borrowed_ideas Anthropic is more bitter lesson pilled imo

Interesting excerpt from @_sholtodouglas on @mattturck's podcast from five months ago on Anthropic prioritisation on coding over other areas including mathematical intelligence h/t: @borrowed_ideas

@dalibali2 it does add calendar events, incl via browser app, which is legit useful

Completely absurd that Gemini is not able to take actions on google apps.

@deanwball i met computer science professors who haven’t even tried Claude code

Among the Twitter-legible US elites, my sense is that the tide really is turning--many people I deem bellwethers have legitimately updated. But if we have another moment that is even plausibly spinnable as "AI is hitting a wall," the view will return.

@macrocephalopod You rang?

Virtu … not a name you hear much any more …

@MikeIsaac anyone who thought Apple should be investing in their own models was very very silly - they can’t even do a good autocorrect

narrative for a while was "apple is behind on AI" — which was true when siri was a bust after trying in house but now they're in a position where they're not spending the insane amounts of capex the other hyperscalers are AND using their models for siri dumb like a fox?

@MikeIsaac As it should be. Anthropic is Claude, model development, and Claude code.

real question: what products that OpenAI has shipped over the past, lets say three years, still exist today or are continuing to be actively developed and prioritized? in my mind the company is now basically ChatGPT, model development and Codex. but tell me what im missing

@rob_mcrobberson @corsaren @anishmoonka @pangramlabs yea ive found an anti ai negation skill probably tackles 80% of what pangram evaluates

@corsaren @anishmoonka @pangramlabs how would you know if they were false positives or not? i sent them several chapters of my short story that pangram insists was ai generated but didnt hear back

Warm colors increase your heart rate. Cool, washed-out tones lower it. Every remake you’ve watched in the last decade has been deliberately color-graded to flatten that signal. It started in 2000. The Coen Brothers shot O Brother, Where Art Thou? in Mississippi during summer, when everything was, in Joel Coen’s words, “greener than Ireland.” They wanted a dusty Depression-era look. Cinematographer Roger Deakins tried every trick in the book: chemical treatments, lens filters, old darkroom techniques. Nothing worked. So they did something no one had done before: digitally scanned the entire film and recolored it frame by frame. Deakins spent 11 weeks turning lush greens into burnt yellows. No feature film had ever been entirely digitally color graded before. Every major studio adopted the technique within a few years. And then the problems started. Modern film cameras don’t capture what your eyes actually see. They intentionally record flat, grey, washed-out footage to capture as much detail as possible. The plan is for the color team to add vibrant color back in later. But the people doing that work stare at grey footage for weeks. Their eyes adjust. One filmmaker admitted he’d bring saturation up to 120% and feel satisfied, then realized the image still looked desaturated to everyone else. He had to crank it to 200% before it looked normal. That’s just eye fatigue. The color draining also happens on purpose. Muting colors hides bad CGI. If a computer-generated background doesn’t quite match the actors, draining the color smooths over the mismatch. The Lord of the Rings extended editions look flatter than the theatrical cuts for exactly this reason: the added scenes had less polished effects, so they were washed out to cover it. Then streaming made it permanent. Bright colors look messy when video gets compressed for phones and laptops. Dull colors look consistent whether you’re watching on a 75-inch TV or a 6-inch phone screen. So studios color their movies for the smallest screen in the room. Your brain registers the difference even if you can’t name it. Your eyes are wired to perceive warm, rich colors as closer and more immediate. Washed-out tones create emotional distance. When a studio drains color from a scene, they’re dampening the emotional signal the image sends to your brain. Old film stock didn’t have this problem. Kodak and Fuji films had rich, punchy color built into the physical chemistry of the film itself. Each brand had a distinct look you could recognize. Digital cameras capture flat, neutral data by default. Getting that warm, vivid “film look” from digital requires skilled work that costs time and money. Most productions don’t invest enough of either. Modern cameras can capture a wider range of colors than film ever could. The technology has never been better. The choices have never been lazier.

@corsaren it’s amusing they bother to use a skill to remove em dashes while not bothering to adjust the syntax and style

what a fucking pussy. gotta nuke these slop shovelers fr

@atelicinvest not to mention anthropic gets to liberally use fast mode and other juiced up options

All the ceos are screaming at their dev team why can't you my move faster look at Claude code Without understanding of the trade off. Or that you can only get away with this if you're breaking new grounds in a market where customers are willing to put up with anything for magic

A lion does not concern himself with making features actually work. Or fixing critical issues like instructions in claude MD getting randomly ignored. A lion simply commits 629 times a day.

@arpitrage @DKThomp Idk - re: puzzle one, did he not even bring up gdp and growth in Central Europe / Asia vs English speaking countries as an explanation?

Great article by @DKThomp on the accuracy and limits of the “smartphone theory of everything view” His bottom line: impacts of smartphones are very culturally contingent theargumentmag.com/p/against-the-…

@georgi_stanoev @buccocapital @pangramlabs Try more negations (don’t do x y or z) Tbh it is fairly easy to trick pangram once you negate common tells like using too many em dashes, certain sentence structures and syntax, certain words

@buccocapital I did that too, and while I find the output phenomenal, @pangramlabs still detects it as fully AI-written. There is still work to be done, but they are getting very close.

I uploaded like a year of my writing to ChatGPT and Claude and asked them both to generate instructions to write like me ChatGPT output is legitimate trash. Garbage slop that is clearly written by AI Claude? It literally writes just like me. Pretty unbelievable

@greyrockave @buccocapital It’s a lot more meticulous than Claude - trade off is your ask needs to be precise and it’s garbage at formatting and design

@buccocapital What do you mean? Wild as in creative or bat shit crazy?

@ThreeHatsInARow @_Jason_Dean_ if you aren’t willing to do ads (and unclear how effective they would be anyway), consumer is bait (tbf, much easier to say in retrospect altho monetization has long been a question given the cogs)

@_Jason_Dean_ OpenAI seems to be realizing the market is way more in enterprise than consumer, at least currently. But that's a pretty big market, I think

It’s an interesting question whether it’s a bearish sign that OpenAI is shifting their inference capacity to answer prompts that have a better gross margin Seems alright to me

@FangYi11101 for sure altho it’s fairly plausible the deals look different given the very unique history between oai and msft

@virtu101 no insight into the deals but prima facie it’s the cloud providers handling distribution/billing/support so it’s a stretch to say Anthropic is the principal for models served on AWS Bedrock

If this is right, then Anthropic’s actual revenue using OAI’s methodology is under $10B (assuming same 80/20 split with cloud providers, and ~30% direct revenue). Under GAAP revenue recognition rules, OAI’s net approach is likely more defensible since the test is who controls the service before it reaches the customer (here the cloud provider rather than model provider).

@EthanChoi7 @OisinO @OpenAI @AnthropicAI Underlying deal dynamics can totally influence revenue rec (principal vs agent)

@OisinO @OpenAI @AnthropicAI Nope. No one asking me to do this. Just laying out what should be straightforward and clear accounting and math end of the day. Don’t know how SEC going to allow different accounting treatment

The way @OpenAI and @AnthropicAI account for revenue / ARR is apples to oranges. Should Anthropic treat their revenue from AWS and other hyperscalers the same as OAI, they would be a materially lower in rev… If they both IPO in the coming quarters, not sure how the SEC is going to let these two companies have different accounting treatment for essentially the same type of revenue. OpenAI TAKES OUT the 80% revenue share that goes to @Microsoft Azure and others so reports this 3rd party revenue on a NET basis in their total revenue. Anthropic INCLUDES the revenue share that goes to @amazon AWS and others in their revenue so reports this 3rd party revenue on a GROSS basis in their total revenue. IMO, OpenAI taking more conservative approach that reflects the reality of the economics of these hyperscaler partnerships.

@kr0der Claude max and pay $20 for gpt plus and have codex do reviews

i’m currently testing Opus 4.6 in Claude Code after using Codex for 3 months straight both have a big downside and i’m currently weighing which one actually slows me down more: Codex - random abstractions, random helpers, not that good at frontend, variable/function naming could use some help Claude - less thorough so it often misses places when implementing features, and produces slightly more bugs in general