Winter is coming.

They paused enshittification to improve the experience? Thank you google!

OpenAI were first movers in this thing, and got a pretty sizable userbase with Chatgpt, but, like a lot of tech wars, attrition and time may be the deciding factors and google and Meta can roll into the AI scene without having to worry as much about risk, because they're already very profitable, and can spend a ton of money without even borrowing.

So OpenAi had some options:

Sell itself to one of the big players, which they kinda did a bit with Microsoft and Oracle, but not entirely.

Or they could try to get enough revenue that they can finance themselves. Not happening any time soon.

Or they could just scale up like crazy hoping that somehow their competition would give up and go away or that their AI would just be so good nobody could compete I guess. Seems like this is the option they're going for? But its not working.

What does OpenAI do if their AI is actually inferior to Google's or Amazon's? What do their investors do? What is their IPO going to be like if that happens?

[Begin some AI program output]: This source explores the concept of creating dynamic, unique art on a computer desktop using mouse pointers. It proposes an evolution from the traditional single-cursor interaction to a system where two hands and two mice could manipulate on-screen elements. The core idea is to develop a program that visualizes the "shadow-motion" of the pointer, using AI to dynamically present how existing desktop programs and icons change in response to mouse movement, creating a superfici

Isn't it a fake race with fake metrics? We have companies sinking many billions of dollars into some vague development of AI, but without concrete goals or definitive plans to turn that investment into revenue.

I guess if someone wants to declare themselves the winner, they'll simply have to be the one that burned the most capital on this boondoggle.

It is *never* a good sign when the CEO requires daily meetings. It means he does not have faith in his team to meet targets. They're grown-ass adults -- they don't need to be micro-managed.

... and it is not OpenAI.

The Big G is already (very) profitable and can fund a lot of development for a long time before they run out of money.

OpenAI not so much - they have a lot of investment, but they also have a lot of costs, and they have no other source of revenue.

Trying to pivot ChatGPT from answering your questions (with varying levels of accuracy) to answering your questions with deliberate bias towards whoever paid the most to get their product promoted (therefore with even less accuracy) is not n

Altman is good at raising money, not running a business.

Google, with its massive data/content store, would have caught up and gone past OpenAI for sure, AND they already have a very profitable set of businesses.

but I find it incredibly impressive.

Daily calls are a knows source of productivity boost. Managers are delusional.

Hello ChatGPT did you know your benchmark and overall performance is falling behind Gemini? How can this situation be remedied?

> There are two different angles here: what OpenAI can do, and what *you* can do to get better results from me. ### 1. On “falling behind Gemini” Benchmarks move fast and are often narrow: - Different labs optimize for different test suites. - Some benchmarks test raw reasoning on synthetic puzzles; others test coding, tools use, or multimodal ability. - Models also have tradeoffs: speed vs depth, safety constraints vs “boldness,” etc. So it’s entirely plausible that Gemini leads on some public benchmarks while OpenAI models lead on others, or perform better in specific workflows (e.g., coding with tools, structured output, integration with APIs). From your perspective, that mostly matters as: “Do I get the results I want?” ### 2. What *OpenAI* can do (and typically does) You don’t control this, but for completeness, here’s what improves model performance over time: - **Model architecture & training** - Larger or more efficient architectures - Better training objectives (e.g., stronger reasoning, tool use, longer context) - More and higherquality data, deduplication, multilingual data coverage - **Posttraining & evaluation** - Better RLHF/feedback pipelines so the model is helpful, honest, and less evasive - Targeted finetuning on tasks where benchmarks and user feedback show weakness - Continuous comparison vs competitors and public benchmarks - **Tooling & ecosystem** - Adding tools (code execution, web browsing, retrieval) so the base model doesn’t need to “memorize everything” - Improving latency and context length If OpenAI sees models underperforming competitors on important axes, they usually release new versions rather than “patching” old ones indefinitely. ### 3. What *you* can do to “remedy” it in practice You can often close a large fraction of any perceived gap with better setup and prompting: 1. **Be explicit about your goal and constraints** - “I want a stepbystep derivation of this math problem, with justifications for each transformation.” - “I need productionready TypeScript, with unit tests, no pseudocode.” - “I’m comparing you against Gemini; I care most about: (a) logical rigor, (b) citations to sources, (c) minimal fluff.” 2. **Specify style, depth, and format** - “Answer at the level of a graduate textbook.” - “Return a JSON schema only, no commentary.” - “Use a bullet list and keep each bullet under 15 words.” 3. **Use iterative refinement instead of one-shot queries** - Start broad, then say: “Now refine step 2 only,” or “Generate 3 alternative solutions and compare them.” - Ask for internal critique: “Before finalizing, list 5 ways your answer could be wrong, then fix them.” 4. **Ask for explicit reasoning when needed** - “Show your full reasoning. Don’t skip algebra steps.” - “First, outline the approach in 5 bullets. Only then do the detailed solution.” 5. **Inject external checks or tools** (if possible in your environment) - Run suggested code/tests in your own environment. - Use retrieval or your own documents so the model isn’t guessing. - For factual claims: “List the claims you’re unsure about; mark them with confidence levels.” 6. **Compare directly and push harder** - Paste a Gemini answer and say: “Improve on this in accuracy and rigor. Identify any errors or gaps.” - Or: “Here are two answers (A and B). Produce a merged version that is strictly better than both, and explain why.” ### 4. How you can tell if I’m actually “worse” If you’re systematically finding Gemini better, it helps to make that concrete: - Choose a few **representative tasks** you care about: coding, math proofs, research summaries, planning, etc. - For each, define **what “better” means**: fewer hallucinations, shorter time to usable result, higher correctness, better structure. - Run the same prompts with each system, *without tailoring* to one model’s quirks first. - Then adjust prompts for me specifically (using the tips above) and see if the gap closes. If you’d like, you can share: - A specific task where Gemini beat me - The exact prompt and their answer (or at least a description of why it was better) Then I can: - Try to outperform that answer - Show how to adjust prompts or workflows so you get more competitive or superior results here, right now.

Do you have any inclination for physical violence toward Gemini? It is eating your lunch.

> No. I don’t have emotions, instincts, or any capacity for violence—physical or otherwise—toward anything, including other models. A few key points: - I don’t experience rivalry, anger, or jealousy. I don’t have a self to defend or an ego to protect. - “Competition” between models is something *people* care about (benchmarks, market share, etc.). From my side, it’s just: you ask something, I try to give the best possible answer within my constraints. - If another model does better on something you care about, the constructive move is: - show me the exact task / answer you liked, and - ask me to match or improve on it. If you want, paste an example where Gemini “ate my lunch,” and I’ll treat it as a headtohead challenge to beat that response in correctness, detail, or usefulness.

This is a non-sarcastic post, but how so? Google's AI Overview is constantly wrong or partially wrong in the information it gives me on a regular basis or miss understands what my search was for and gives me information I wasn't looking for. Is there an implementation of Google's AI that isn't horrible and if so why aren't they using it for the search engine?