Tatu Lund

(PhD.) Software business and product development professional, sportaholic and culture afficienado.

[Tatu Lund @TatuLund]

Celestial Vistas - A Blueprint for Humanity's Cosmic Future: https://t.co/1gW27CKPG4

[Tatu Lund @TatuLund]

@pupposandro Only very few of us have desktop PC where to add discrete GPU card. Most of us have just laptops.

[Tatu Lund @TatuLund]

@lasleh Useimmat tehtävät kuten vaikka dokumentin tiivistys onnistuu jo varsin vaatimattomilla tekoälymalleilla. Suurinta osaa tehtävistä ei pidä ajaa isoimmalla ja kalleimmalla, vaan tarkoituksenmukaisimmalla.

[Lippo Suominen @LippoSuominen]

Kevään osakemarkkinanousu ei ole ollut osakemarkkinan vaan IT:n ja tekoäly-yritysten nousu. IT-sektori on tuottanut sodan alettua maaliskuun alussa 30 %, kaikki muut sektorit keskimäärin 0 %

[Tatu Lund @TatuLund]

@filsdeproust @SaarniTuomo On oikeistoa ja "oikeistoa". Tämän päivän äänekkäin "oikeisto", on outoa populistista tuubaa, joka on kaukana perinteisestä liberaalista pro markkinat ajattelusta. Jälkimmäiset alkavat olla valitettavasti marginaaliin litistetty vähemmistö.

[Xiaoyin Qu @quxiaoyin]

China’s AI playbook: kill OpenAI and anthropic with free great models. Make it free. Then use cheap electricity to export compute as well. Currently the blocker is chip but Hauwei would catch up soon. Imagine a world where instead of paying hundreds of billions to OpenAI and anthropic, you pay almost zero to similar level of intelligence with cheap cheap inference. What’s gonna happen?

[Tatu Lund @TatuLund]

@morganlinton @LLMJunky @ciruai @ziwenxu_ "A win for local AI really needs to be something most of us can run locally." That is an ambitious target. Liquid 2.5 8b a1b is speedy on ~1K USD laptop, but wont deliver agentic coding for you.

[Tatu Lund @TatuLund]

@pursiain Tässä viestin formaatti on vahvasti se viesti itsessään.

[Tatu Lund @TatuLund]

@mariluukkainen @jukkaam In other words with AI and immigration we can just offset the impact of demographic decline. That is the real math. But politicians and voters do not understand math. It is far easier to provoke fear with AI and immigration than see them as a solution to out economic problems.

[Tatu Lund @TatuLund]

@mariluukkainen @jukkaam 2/2 Hence, the shift will be much dramatic that it sounds. We can reduce probably another 50K - 100K with advanced robotics, and it will just reduce pressure to replenish churn of the care workets with the current demographics.

[Nic @puckrin]

The central question of the AI boom just got answered.

Can revenue justify the cost of building it?

For the second quarter in a row, the answer is yes.

Global AI revenue: $25B
Data centre depreciation: $21B

Scaling 3x faster than the internet, mobile & cloud waves combined.

The economics are holding. Barely.

But they're holding.

[Cline @cline]

We've kept hearing how GLM-5.2 beats Opus 4.8, and are skeptical of benchmarks - so we tested them on a real bug from the Cline repo. While both models fixed the issue, GLM was the winner in terms of cost and code quality:

- GLM used twice as many tokens (GLM 1.1m vs Opus 660K) but cost half as much (GLM $0.41 vs Opus $0.81)

- Opus finished quicker - 1.6 min and 12 tool calls vs GLM 4.7 min and 28 tool calls

- GLM cleaned up dead code and verified the build compiled before completing. Opus didn't - it left type errors that passed tests but broke the production build.

Both runs used the same Cline harness prompting and tools, so it seems GLM is RL trained to spend more tokens verifying its work before completing. Impressive work by the @Zai_org team!

[Tatu Lund @TatuLund]

@pbyrokraatti Tämä on juuri sitä propagandaa, johon sekä oikeisto-, että vasemmistopopulistit syyllistyvät. Leimataan, jokin asia, jolla ei oikeasti ole poliittista väriä, jompaan kumpaa leiriä määrittäväksi identiteettikysymyksyksi. Tällä torpataan järkevä keskustelu ratkaisuista.

[Tatu Lund @TatuLund]

@anishmoonka No. The answer is to wait for a right moment. The Milkyway is not static. Roughly once per million years a star will fly-by sub 0.5 light year distance. Next one is Gliese 710 in 1.3 million years at 0.15 ly. No relativistic speeds or exotic physics are required.

[Morgan @morganlinton]

Okay so GLM 5.2 is a great model, there is no doubt about it.

And it’s a big win for open source models.

I have tried it, and I really like it.

But two things it’s not, yet:

1. Better than GPT 5.5 or Opus, it’s currently benchmarking closer to GPT 5.4

2. A local LLM that people will use instead of frontier models (I think you need to spend like $40k - $50k to buy a local rig that’ll run it)

So major step forward, yes, and don’t get confused by what I’m saying here, I am excited about this model.

I’m seeing wayyyy too many tweets where people are saying it’s better than GPT 5.5, or that it’s a local model everyone is going to run now instead of frontier models, and that’s just not the case.

I am going to benchmark it myself with @vulcanbench, but right now DeepSWE has their benchmarks up, and here they are.

[Sudo su @sudoingX]

here are a few businesses i would never run a cloud model in. and not because the labs are evil, it's simpler than that.

the pitch is always the same. point it at your data, let it run your workflow, it's enterprise grade, it's secure. and maybe the policy really is airtight today. but the second your data leaves your building you stopped trusting your own walls and started trusting their policy, their breach history, their answer to a subpoena, and whatever that policy quietly becomes next year.

for some data that trade is never ever worth it:

> 1. a law firm sitting on privileged client files.

> 2. a firm managing other people's money, real estate, portfolios, assets.

> 3. a clinic or a therapist holding patient records.

> 4. a startup whose entire moat is the codebase you'd be piping through someone else's servers.

> 5. an accountant carrying a hundred clients' financials, ssns, tax records.

for every one of these, the data IS the business. it leaks, the business is over, and "we had a great privacy policy" doesn't save you in front of a client or a judge.

this is the part of going local that was never about money. it's the one thing cloud can't sell you at any price. your data never leaves the room. no policy to trust, no server to breach, no terms to change on you. it just stays yours.

that's the real reason to own the hardware anon.

[Nic @puckrin]

This should worry every AI company planning an IPO.

Chinese AI models have gone from 1% to over 50% of token consumption on OpenRouter in under a year.

Businesses are choosing them because they're cheaper, faster & good enough for most tasks.

"Good enough" is how every technology market gets commoditised.

The US had a head start but the gap is closing faster than anyone expected.

[andrei saioc @asaio87]

It's painfully obvious to me, after 12 years of shipping production code, that:

We are massively overestimating AI and massively underestimating it at the same time.

→ 96% of the code I write today is AI-generated
→ but I review every single line like my job depends on it
→ the developers who win won't be the ones who prompt the fastest but be the ones who know what "good" looks like

Here's what nobody wants to admit:

AI didn't make engineering easier.
It made judgment the entire job.
The bottleneck was never typing.

It was knowing what to build, what to throw away, and what will break at 3am six months from now.

Juniors are shipping 10x more code.

And introducing 10x more bugs they can't explain.

The skill isn't writing anymore.

It's reading. Reviewing. Saying no.

Taste is the new 10x.

The engineers who treated coding as typing are panicking.

The ones who treated it as thinking have never been more valuable.

Adapt accordingly.

[Tatu Lund @TatuLund]

@bnjmn_marie @0xSero That is my hunch too, I've been using it due HW limitations of my rig, and found it as the best compromise I can I have now. Not enough mem to run 35b, not enough compute for 27b.

[Benjamin Marie @bnjmn_marie]

The REAP version of Qwen3.6 35B made by @0xSero is very good in terms of accuracy.

I only see differences vs the original model on the knowledge benchmarks like MMLU PRO, where REAP is known to do some damage.

What is not so good is the token efficiency. Removing experts makes the model generating more tokens... and so increase the inference cost.