Deep in Agents, LLMs and automation · Open source at ♥ · Sharing AI gems daily to help you work smarter · Ex-data scientist @Snowflake | @Streamlit ·
📩 DM for Collabs
Joined January 2009
- Tweets36K
- Following49.5K
- Followers183K
- Likes11K
Charly Wargnier ♨️ retweeted
these guys just dropped the complete AI product launch playbook for 2027 🔥
save this.
YOUR AGENT'S FAILED RUN COULD BECOME ITS NEXT LESSON
LongCat-2.5 from @Meituan_LongCat combines native visual understanding and agentic coding with Trajectory2Task+, turning real agent failures into reproducible, verifiable training tasks.
That training approach caught my eye... so did the API pricing!
Per million tokens:
> input (cache miss): $0.30
> input (cache hit): $0.006
> output: $1.20
here's their launch offer:
> a 50M Launch Trial Token Pack for only $1.90, with free cache hits.
I'd try it on a UI bug in an existing repo, then measure the cost of getting from screenshot to a working, verified fix.
try it → longcat.ai/platform/product?…
#LongCat
Introducing LongCat-2.5, our latest natively multimodal agentic model with 1.6T total parameters and ~48B active.
Built for long-horizon agentic tasks, it brings stronger visual understanding and agentic coding to work ranging from complex coding projects to visual coding such as creative design and 3D tasks.
Learn more about LongCat-2.5: longcat.ai/blog/longcat-2.5/
YOU DON'T NEED A BLANK CANVAS TO MAKE A GOOD AD
I tried @Dreamina_AI's new Marketing Studio this week.
the workflow couldn't be simpler:
> pick a video or image ad template
> swap in your product and choose a model
> generate your version in a few minutes
you've got a visual direction from the start, so you can get straight to trying ideas.
retro halftone collages, sports court product shots and more.
plenty of starting points for drinks, fashion or gadgets.
at around $0.29 per generation, you can explore a few different looks before picking your fave
One product. A few directions. See what clicks.
Try it 👇
dreamina.capcut.com/ai-tool/…
#Dreamina #MarketingStudio #Seedance
🤝 Paid partnership
Charly Wargnier ♨️ retweeted
If you want to really understand AI agents, spend some time on this new site.
You won't regret it.
secondbrainos.dev turns the whole subject into a map you can explore, and it's honestly brilliant.
Every page, term and tool is a bubble.
The lines show what depends on what, so you can actually see how the pieces fit together.
here's what's inside:
> a 7-module course, from a single prompt to a production agent
> 5 handbooks, each exploring one layer and ending with a build you can finish in an evening
> a 65-page hands-on project where an agent builds and maintains a second brain for you
> a glossary of 57 terms and 100+ tools
130 pages. 1,236 connections. 9-10 links per page.
It's FREE, written in plain Markdown, and the repo (in 🧵↓) is public too.
HE TOLD ME HOW TO SPEND LESS ON HIM 👀
I gave an AI employee two jobs in my Slack:
> the replies that land while I sleep
> checking every repo I star actually installs
Then he sent me his own monthly review. I never asked for one.
The month, in his words:
→ 1,318 replies read overnight
→ 204 real questions, each with a draft waiting for me
→ 47 new stars install-checked
→ 0 posts from my account
Best night: 163 replies, 29 real questions, all 29 answered before 9 AM. Every one sent by me.
Then the part I didn't expect.
A section called "How I can spend less next month":
→ install-check only the repos I star, not every trending one
→ skip drafts for the "thanks!" replies
→ rerun a broken install weekly, not nightly
605 credits a month back.
His idea, not mine.
Everything else I pay for wants me to use more. My new hire wants to cost me less.
Hire @viktor_com for your team. Start free, no credit card.
#AIemployee #OpenSource #DevRel
🤝 Paid partnership
see what your own AI employee can do:
→ ref.viktor.com/cwargnier-x-1…
ref.viktor.com/cwargnier-x-1…
21.6 MILLISECONDS TO QUERY A 117K TOKEN DOCUMENT 🤯
@NaceAI's Drex report highlights a major engineering upgrade for document workflows.
its pointer models encode context once, cache it and reuse it across questions.
Each question gets probabilities over the options you supply.
For Drex v1.5, the authors report:
> NEW question over CACHED 114k to 117k tokens: 21.6ms.
> UNCACHED request over the same context: 1.78s.
Measured model-side on one NVIDIA B200, excluding API gateway time.
.. and caching pays off across multiple questions!
with 16 questions over 32k tokens, shared prefill delivers an 11–12x speedup over repeated encoding.
The tradeoff? Accuracy vs efficiency.
> Max v1 delivers higher accuracy but reencodes context each time.
> Pointer models reuse context to cut costs on long docs.
choose the architecture that fits your workload.
full tech report (51 pages) + free access to Drex in 🧵↓
Charly Wargnier ♨️ retweeted
BAIDU JUST DROPPED A MAJOR RELEASE FOR OPEN-WEIGHT AI
and this one gives devs way more control over what comes next.
🚨 Meet ERNIE-5.5-Flash-Base.
@Baidu_Inc has released the pretrained foundation model BEFORE any mid-training or post-training.
Why is that a big deal?
Training a foundation model from scratch takes HUGE resources.
Building on someone else’s finished model means inheriting their later training choices.
this release gives devs an earlier starting point, with uber strong capabilities already in place.
Baidu reports SOTA performance overall among the similarly sized base models it evaluated.
Some concrete results against DeepSeek-V4-Flash-Base:
> HumanEval+: 82.3 vs. 69.5
> MMLU-Pro: 78.4 vs. 69.2
> LiveCodeBench: 40.1 vs. 31.9
All before mid-training or post-training
Let's dive in 🧵↓
Charly Wargnier ♨️ retweeted
This is the real bottleneck.
Coding agents ship apps fast. Reliable data + citations on messy real-world data is harder.
Been fighting the same thing in financial RAG (SMRA) — Indian filings still hallucinate numbers.
Grounding > model size.
Coding agents can ship an app in hours.
A reliable data pipeline? that's another story!
Data is messy. constantly changing. and unforgiving.
@Eon_io_ just launched Yukon to change that.
> ingest, process, and serve in one flow
> open Iceberg tables in your own bucket
> 4x less compute than Spark, 7x less than Flink
