Tom Keldenich

609 posts

Tom Keldenich

Tom Keldenich

@tokelde

Building customer-facing chatbot for a french bank in Paris

Paris Katılım Mayıs 2024
141 Takip Edilen175 Takipçiler
Tom Keldenich retweetledi
Reid Wiseman
Reid Wiseman@astro_reid·
Only one chance in this lifetime… Like watching sunset at the beach from the most foreign seat in the cosmos, I couldn’t resist a cell phone video of Earthset. You can hear the shutter on the Nikon as @Astro_Christina is hammering away on 3-shot brackets and capturing those exceptional Earthset photos through the 400mm lens. @AstroVicGlover was in window 3 watching with @Astro_Jeremy next to him. I could barely see the Moon through the docking hatch window but the iPhone was the perfect size to catch the view…this is uncropped, uncut with 8x zoom which is quite comparable to the view of the human eye. Enjoy.
English
4.1K
45.2K
263.4K
18.1M
Tom Keldenich
Tom Keldenich@tokelde·
@svpino Not surprised, small LLM have trouble with structured outputs (even on Gemini/Claude tier, let alone SLM)
English
1
0
0
608
Santiago
Santiago@svpino·
I'm running Gemma 4 on my computer with Ollama. Unusable with Claude Code. It can't even load and execute skills, so I had to stop. But the model is pretty decent as a chatbot using the Ollama UI. I've been cross-posting questions across Claude and Gemma 4, and I can use Gemma's answers without any problems. I wish we had a better UI harness for the model (with projects, memory, etc.)
English
221
17
619
158.5K
Lisan al Gaib
Lisan al Gaib@scaling01·
ANTHROPIC HAD MYTHOS INTERNALLY SINCE FEB 24
Lisan al Gaib tweet media
English
84
88
2.5K
1.1M
Tom Keldenich
Tom Keldenich@tokelde·
@qalfredoai Just finished the firsts 7 levels It could make a great mobile app game
Tom Keldenich tweet media
English
0
0
0
318
Alfredo
Alfredo@qalfredoai·
Just spent 10 minutes playing the ARC-AGI-3 games and i genuinely cannot get over it. You figure out the rules yourself in like 2-3 minutes. no instructions. just vibes. GPT-5, Gemini 3 and Claude score below 1% on these. Try it yourself: arcprize.org/arc-agi/3
English
9
10
129
23.6K
Tom Keldenich
Tom Keldenich@tokelde·
@verge I guess the best form for AI video app is simply what have been done since Midjourney: a gallery of the best generations of the day
English
0
0
0
335
Tom Keldenich
Tom Keldenich@tokelde·
@EthanHe_42 Tried it the other day to transform a pain au chocolat into a turtle and was pretty amazed with the result !
English
0
0
0
95
Ethan He
Ethan He@EthanHe_42·
couldn't tell this was AI until the very end
English
94
44
691
58.6K
Ashwin Hegde
Ashwin Hegde@ashwinhegde19·
what are you currently working on at OpenAI ? @steipete
English
1
0
33
8.1K
Tom Keldenich
Tom Keldenich@tokelde·
Reminder if prev. director of AI @Tesla debate their ideas with LLMs, it’s probably a habit you should copy for yourself
Andrej Karpathy@karpathy

- Drafted a blog post - Used an LLM to meticulously improve the argument over 4 hours. - Wow, feeling great, it’s so convincing! - Fun idea let’s ask it to argue the opposite. - LLM demolishes the entire argument and convinces me that the opposite is in fact true. - lol The LLMs may elicit an opinion when asked but are extremely competent in arguing almost any direction. This is actually super useful as a tool for forming your own opinions, just make sure to ask different directions and be careful with the sycophancy.

English
0
0
1
42
Tom Keldenich
Tom Keldenich@tokelde·
Hooks in codex are basically middleware for AI agents They let you inject your own scripts directly into the agent loop at key moments like: • before/after tool use • when a user submits a prompt • when a turn finishes What this unlocks: → Log conversations to your analytics stack → Block secrets (like accidentally pasted API keys) → Auto-summarize chats into memory → Enforce team rules / guardrails → Customize behavior per repo or directory The key idea is that instead of just prompting the model, you’re also programming the loop around it.
English
0
0
0
45
Tom Keldenich
Tom Keldenich@tokelde·
@TheVixhal Interesting to see how it would works, but if you have a lot of documents, you'll blow up your context window
English
0
0
0
67
Tom Keldenich
Tom Keldenich@tokelde·
When building LLM app, latency can quickly become a bottleneck. Here’s my approach to reduce it by ~20% : I set the thinking level to minimal. Today’s LLM have thinking mode activated by default. It makes the model smarter, so it can accomplish harder task. But thinking makes LLM slower, and this might not align with your app expectation. Thinking is powerful, but in many cases, you don’t need it at all. Raw frontier models can do simple tasks easily : summarizing, formatting text, generating documentation, etc. With these type of tasks, you don’t need extra reasoning tokens. You just need fast generation. If speed should be on your side, and your task is simple, you can deactivated thinking in your model config. Here’s how to do it with Gemini :
Tom Keldenich tweet media
English
0
0
0
26
Tom Keldenich
Tom Keldenich@tokelde·
@LukeParkerDev My very simple workflow to generate code that works is to first have a conversation with the llm, however long it might be, then ask it to implement it, with all the relevant code needed (The more, the better)
English
0
0
0
133
Luke Parker
Luke Parker@LukeParkerDev·
im am this close to crashing out. every AI just does dumb stuff unless you are so specific you may as well code. it can help for mass migrations once you've already done the shape and exact impl, and have a bunch of boring work. im so sick of trying to wrangle it lol
English
151
36
966
45.8K
Tom Keldenich
Tom Keldenich@tokelde·
@sweatystartup AI isn't a bubble and creating agents is the most valuable skill you can learn right now
English
0
0
0
20
Nick Huber
Nick Huber@sweatystartup·
AI about to get 20x expensive. These $200 / month claude subscriptions are burning $5,000 worth of credits. The bubble is going to pop and it will pop soon.
English
397
153
2.5K
231.6K

Keşfet

@Astro_Christina @AstroVicGlover @Astro_Jeremy @svpino @scaling01 @qalfredoai @verge @grok