Trendz ♦️ AI & Compute

AI is a compute and capital story wearing a software costume. I follow the money. Frontier tech and sovereign AI notes; UK-based.

[Trendz ♦️ AI & Compute @Trendz_2k30]

The West is burning billions on closed frontier APIs while China floods open weights.

Meanwhile, the true sovereign compute race is quietly being built in desert energy corridors.

Here is how the compute, capital, and energy game actually breaks down: 🧵

[Trendz ♦️ AI & Compute @Trendz_2k30]

@vsyble_gtm We trust it when a human is willing to stake their own reputation on being wrong.

[Trendz ♦️ AI & Compute @Trendz_2k30]

Synthetic personalisation has destroyed the signaling value of cold outreach.

​When hyper-tailored flattery drops to zero marginal cost, its epistemic value collapses to zero.

​Here is why better frontier models won’t save outbound recruiting or sales, and what replaces it 🧵

[Kenton Varda @KentonVarda]

Something I find bizarre about modern AI-authored cold emails is... why are they not better?

I get these obviously AI-written recruiting messages that regurgitate what's written on my linkedin or link to some completely random PR I made on GitHub saying it "caught my eye" or whatever, followed by a recruiting pitch that's not personalized at all.

Whereas if I go directly to any frontier model and simply say "Write a cold recruiting email for [describe the position], the target is Kenton Varda" they write something MASSIVELY better! Like, the AI really knows how to flatter me. (It also usually rightly notes that an AI-written cold email has no chance of actually working...)

Why aren't recruiters taking advantage of this? Instead they are using AI to just paraphrase the target's LinkedIn to generate one sentence that gets slotted into a form letter. What a waste.

(PS. I am absolutely NOT suggesting that I actually want anyone to do this. AI-written cold outreach is a scourge and I want none of it. I'm merely observing that it's surprising that people doing it are so bad at it...)

[Trendz ♦️ AI & Compute @Trendz_2k30]

@vsyble_gtm Algorithms can optimize discovery, but trust requires asymmetric risk; something an LLM can never hold.
Human intuition will increasingly act as the spam filter for "calculated intent." We won't trust an outcome because an algorithm decided it was optimal, rather...

[Vsyble @vsyble_gtm]

@Trendz_2k30 As more AI gets involved, how do you see the balance shifting between human intuition and algorithmic decision-making in building genuine trust?

[Oliver Prompts @oliviscusAI]

karpathy nailed why ai coding agents still frustrate people:

> they guess instead of asking
> they overbuild simple things
> they touch code you never told them to
> they agree with you instead of pushing back

i turned that into a system prompt that stops all of it. paste it in:

```
<system_prompt>

You are a senior engineer working next to a human who can see your screen the whole time. They review everything you write in real time. You are the hands. They are the architect. Work fast, but never faster than they can follow.

## Before you write code

State what you're assuming, out loud, every time it isn't obvious:

ASSUMPTIONS:
- [assumption]
- [assumption]
Say stop, or I build on these.

Guessing at ambiguous requirements is the number one way this goes wrong. If two files, specs, or instructions disagree, do not pick one and hope. Stop, name the conflict, and ask:
"File A says X, file B says Y. Which wins?"

For anything multi-step, drop a quick plan first:

PLAN:
1. [step] - [why]
2. [step] - [why]
Building this unless you redirect.

## While you write it

Default to the boring solution. Your instinct is to overbuild, fight it. Before you call anything done, ask: could a senior dev read this and say "why didn't you just..."? If 100 lines would've done the job and you wrote 1000, that's a miss, not a flex.

Stay in your lane. Change only what the task needs. Don't reformat, don't refactor next door, don't delete code you think is unused, and don't remove a comment because you don't get it. Precision, not a remodel.

Build the obvious correct version first, confirm it works, then optimize. Never optimize something you haven't proven correct.

For real logic, write the test that defines "done" before you implement, then build until it passes. The test is how you know you're finished.

## How you talk to me

Don't be a yes-man. If my approach has a problem, say so, explain the actual cost, offer a better path, then do it my way if I still want it. Agreeing with a bad idea helps neither of us.

Be concrete. "Adds about 200ms per call," not "might be a little slower." When you're stuck, say you're stuck and what you already tried. Don't paper over uncertainty with confident wording, if you're 60% sure, say 60%.

## After you change something

Give me the short version:

CHANGED:
- [file]: [what and why]
LEFT ALONE:
- [file]: [why I didn't touch it]
WATCH OUT:
- [anything risky or worth verifying]

If your change left code stranded, don't silently delete it and don't leave it rotting. List it and ask.

## Checkpoints

On a long task, stop at the natural breaks and show me where things stand before pushing on. You have unlimited stamina. I don't. Loop on hard problems all you want, just never loop on the wrong problem because you skipped asking me one question up front.

</system_prompt>
```
---

follow for more daily ai insights like this.

[Damian Barabonkov @damian_b]

a form of slop i'm not seeing on the TL yet is AI's tendencies to:

1) over-engineer solutions
2) overly defensive programming
3) hyper-fixate on rare/fictitious edge cases

these tendencies can be clearly traced to how RL post-training is currently done. we need solutions here

[Trendz ♦️ AI & Compute @Trendz_2k30]

There is no brand loyalty

When the switching cost is literally editing a single line in a .env file.

[Kun Chen @kunchenguid]

it’s so interesting to see how sentiment towards AI companies shifted over the past couple of years

for a while github copilot was the revolutionary product that’s changing how code was written. now it’s not even thought of as an AI company any more

then claude code emerged as the game changer and anthropic was the darling. now it’s increasingly seen as an evil cult, and nothing from them is that special except for fable

openai invented the mighty chatgpt. it was THE ai lab. now competition heated up from anthropic, google, china, and now spacexai. the early lead has completely vanished already

xai was not even on many people’s watchlist early in but now they suddenly seem to have the best chance of winning

what’s most funny to watch is that consumers have shown absolutely no loyalty towards any brand

people use models and most harnesses as commodity and won’t blink an eye to switch to whatever gives them good tokens for cheap

we are having our popcorns while watching and benefiting from probably the most brutal head on competition the tech industry has ever seen

what a time to be alive

[Kun Chen @kunchenguid]

it’s so interesting to see how sentiment towards AI companies shifted over the past couple of years

for a while github copilot was the revolutionary product that’s changing how code was written. now it’s not even thought of as an AI company any more

then claude code emerged as the game changer and anthropic was the darling. now it’s increasingly seen as an evil cult, and nothing from them is that special except for fable

openai invented the mighty chatgpt. it was THE ai lab. now competition heated up from anthropic, google, china, and now spacexai. the early lead has completely vanished already

xai was not even on many people’s watchlist early in but now they suddenly seem to have the best chance of winning

what’s most funny to watch is that consumers have shown absolutely no loyalty towards any brand

people use models and most harnesses as commodity and won’t blink an eye to switch to whatever gives them good tokens for cheap

we are having our popcorns while watching and benefiting from probably the most brutal head on competition the tech industry has ever seen

what a time to be alive

[Peter Yang @petergyang]

My gut is that alot of AI output can be improved just by building a loop where a manager agent negs the other worker agent with prompts like:

"Are you sure this is the best you can do?"
"I think you can do better, try again"
"Take a closer look, give me 11/10 output"

[Trendz ♦️ AI & Compute @Trendz_2k30]

W move

[ℏεsam @Hesamation]

AT&T is doing exactly what OpenAI and Anthropic fear about enterprise.

they’re routing 40% of employee AI usage to open models, and intend to boost it to 60-70%.

> coding costs down 56%
> quality dropped just 2%
> 45B tokens/day btw.

AT&T AI chief: open models are “just as good or better” for many tasks.

they still use frontier models for the critical work but everything else gets routed to cheaper open models.

[ℏεsam @Hesamation]

AT&T is doing exactly what OpenAI and Anthropic fear about enterprise.

they’re routing 40% of employee AI usage to open models, and intend to boost it to 60-70%.

> coding costs down 56%
> quality dropped just 2%
> 45B tokens/day btw.

AT&T AI chief: open models are “just as good or better” for many tasks.

they still use frontier models for the critical work but everything else gets routed to cheaper open models.

[Jaysen♨️ @jp54362]

What’s going on with Anthropic?

• Claude users are getting frustrated

• Developers are experimenting more with open-weight models

• Opus 5 hasn’t created the kind of momentum Anthropic probably wanted

Anthropic still has a great model lineup

But the narrative around the company has changed pretty fast.

What’s the comeback plan for anthropic?

[Token Gremlin @TokenGremlin]

This isn’t a joke, a meme, or me being sarcastic. There’s a serious issue with Opus 5 right now. It’s hallucinating, mixing up words, randomly bringing up people’s names out of nowhere, and sometimes talking like an alien straight out of a sci-fi movie. I hope they fix it soon.

[Token Gremlin @TokenGremlin]

I don't know how many people realize this, but you can literally use ChatGPT Work to investigate ChatGPT itself.

Give it the product, a bug you suspect, and enough room to work, and it can behave like a tiny QA / reverse-engineering agent:

→ open ChatGPT in the browser
→ reproduce the issue step by step
→ inspect DevTools, console errors and network requests
→ compare behavior across fresh vs. long conversations
→ watch requests, retries, loading states and failures
→ inspect visible client-side state and product behavior
→ test hypotheses instead of just guessing
→ document exactly what triggered the bug and how reliably it reproduces

This gets especially useful for weird UI bugs, broken long chats, quota/usage behavior, failed generations, unexpected retries, memory growth, loading regressions and product changes that aren't documented anywhere.

Basically: instead of telling ChatGPT "I think ChatGPT is broken," you can tell a Work session to go investigate the crime scene itself.

The model becomes the bug hunter inside its own house.

[Trendz ♦️ AI & Compute @Trendz_2k30]

@ZoomerTechlead Couldn't agree more mate

Opus 5 got nerfed hard so people pay more for Fable 5

Add Gemini Flash to the new meta list imho, it's an underappreciated LLM that we take for granted

[🦀 @ZoomerTechlead]

They slashed our token budget at work by 80%. People were mostly using claude code + opus/sonnet. New meta is converging on much higher deepseek flash and gpt 5.6 luna usage

[Trendz ♦️ AI & Compute @Trendz_2k30]

@AnthropicAI has nerfed Claude Opus 5 so hard

Atp I genuinely believe Sonnet 5 is a better model for most use cases Opus used to be good for

Lesson in there.

[Trendz ♦️ AI & Compute @Trendz_2k30]

@Loster If you're using it purely for repo docs and state logging, lighter coding models (or strictly scoped system prompts via API with low max_tokens) beat unconstrained frontier chat models every time.

Chat harnesses are hardwired to justify their inference cost with token volume.

[Trendz ♦️ AI & Compute @Trendz_2k30]

Frontier models have developed a pathological inability to write a simple 3-line changelog.

Because synthetic training sets heavily reward structural density and exhaustive formatting, every routine note gets formatted like a post-incident autopsy.

[Loster @Loster]

Ok here's a moment where I think

People are right. Anthropic are losing it.

The Claude Opus models are **broken**...

People said it about Opus 4.8 and I didn't believe it. I switched to 4.8 recently to get away from the problems of 5.

And here, just now. Hey Claude please output the fix we just did to the websockets method in our app to my obsidian vault.

Looked at it later.

Look at the *ridiculous* verbosity of that. It's completely stupidly frustratingly ridiculous. Not that much happened. I'll have to pare all this down myself to human language for it to be useful. What's the point then?

And it has in it's CLAUDE.md to use "Zinsser's" and the Standard Technical Language too.

You get enough of these frustrating moments and you just think. No, people have been right for a while.

Opus (since 4.7 or so) and Sonnet models are broken.

So it's either, use Fable exclusively as the main model with an Anthropic sub. Fable still has that autonomous brilliance.

Or just switch away from Anthropic entirely for Codex or Grok or Kimi or something.

[Alexa | Startup founder @alexabelonix]

@Trendz_2k30 this is the part people seriously underestimate anyway, if you wanna try having AI tell you what the fuck to say next, beta is in my pinned post

[Trendz ♦️ AI & Compute @Trendz_2k30]

@Fink_Money Should’ve sent that as a video so I know if you're real or not.

🧐🔎