Overhead view of a wooden table with open laptops, half-empty mugs, scattered notebooks with hand-drawn model architectures, and a phone showing a Slack thread mid-conversation — warm golden-hour light through tall windows, only hands and tools visible
Free · Open to All AI Builders

Nobody builds
alone.

A free Slack for AI builders — the ones shipping wrappers at 2 a.m., debugging LoRA runs nobody warned them about, and betting on themselves when every recruiter stopped calling.

Join Free — Find Your PeopleKeep the Fire Going →

Scroll to read the story

The Problem

Building AI is the most technically demanding thing most people will ever do — and almost everyone does it alone.

73%

of solo AI builders report going weeks without meaningful technical conversation

State of Independent AI, 2025

61%

considered quitting during their first major model failure with no one to call

Indie Hacker Survey, Q4 2025

2.3x

more likely to abandon a project when building without a peer community

ML Persistence Study, Stanford

From Stack Overflow · Asked 8 months ago · 0 answers

0
votes
0
answers

LoRA fine-tune diverges after 200 steps — loss spikes to NaN. Using LLaMA-3.1-8B. Tried reducing lr from 2e-4 to 5e-5. Nothing. Please, anyone.

0
votes
0
answers

My RAG pipeline returns confident hallucinations on every third query. Context window is 16k. Embeddings look fine in cosine space. 3 weeks on this.

0
votes
0
answers

Quantized model passes eval but fails completely on production inputs. GGUF vs GPTQ both broken. No one at my company knows ML. I might just quit.

Thousands of questions. Silence where answers should be.

"The mass layoff hit in November. I had a half-built fine-tuning pipeline, a runway of four months, and zero people to ask if I was even on the right track. I almost convinced myself the problem was me — that I wasn't smart enough to be doing this without a team around me."

Priya Nair

Former Staff ML Engineer, now building Kestrel AI · Grove member since March 2025

The Turning Point

At 2:14 AM, Priya posted into the dark.
Seventeen strangers answered.

In 33 minutes, a LoRA fine-tune that had been broken for six hours was fixed. Nobody asked for credit. Nobody charged a consulting fee. They just showed up — because that's what Grove is.

#debug-together·Grove Slack Community
live
PN
Priya Nair#debug-together2:14 AM

ok i give up. LoRA fine-tune on llama-3.1-8b keeps diverging at step 200. loss spikes to NaN. been at this 6 hours. anyone awake

JO
Jonas Obermeyer2:17 AM

awake. post your training config. specifically r and lora_alpha

TW
Tomás Wu2:18 AM

also — what's your batch size and are you using gradient checkpointing? NaN at exactly 200 steps is suspicious

PN
Priya Nair2:19 AM

r=64, alpha=128, batch=4, yes gradient checkpointing. here's the config [paste]

AS
Amara Seck2:23 AM

your alpha/r ratio is 2 — try 1:1 (r=64, alpha=64). high alpha inflates gradients early. seen this exact spike before

🔥💡
JO
Jonas Obermeyer2:31 AM

also reduce warmup_ratio to 0.03 if you're below 0.05. at 200 steps with your dataset size you're still in warmup when it spikes

👆
PN
Priya Nair2:47 AM

IT'S TRAINING. step 400 and loss is 1.2 and falling cleanly. i'm crying a little. thank you all. genuinely.

🎉🔥❤️🙌

This thread has 47 replies and has been bookmarked 203 times. Priya's startup raised a pre-seed round six months later.

"I didn't know anyone in that channel. I still don't know what most of them look like. But I know they showed up at 2 AM for a stranger's loss curve — and that changed what I thought was possible."

— Priya Nair

The Scale

This is what happens when builders stop building alone.

Grove has been free since day one. No paid tiers. No paywalled channels. Every number below was made possible by members who chose to support it.

0

Members

AI builders, prompt engineers, fine-tuners

0

Countries

From Lagos to Ljubljana to Louisville

0+

Jobs landed

Through intros made in Grove channels

0

Open-source projects

Born in #build-in-public

Where the work actually happens

Six channels. Thousands of real conversations every week.

#debug-together

Real-time debugging with people who've seen your exact error before

1.2k members

#fine-tuning-lab

LoRA, QLoRA, RLHF, DPO — every technique, every war story

847 members

#build-in-public

Ship updates, get honest feedback, celebrate small wins loudly

1.1k members

#hiring-and-intros

340+ jobs landed through warm intros from people who know your work

934 members

#late-night-builds

For the 2 AM builders. No judgment. Just presence.

623 members

#mentorship-pairs

Retired engineers and new builders matched by domain and timezone

412 members

Who's here

The Solo Founder

Left FAANG with a thesis and a savings account. Building an AI product alone because the equity math finally made sense.

38% of members

The Pivoting Builder

Bootcamp grad or self-taught dev who realized ML ops was where the next decade of work was going. Learning in public, shipping in public.

41% of members

The Kitchen Table Mentor

Retired or semi-retired engineer with 20 years of distributed systems scar tissue. Here because they remember having no one to ask.

21% of members