Steven (Batman) Batchelor-Manning
@S_BatMan

Agent Shepherd | Founder | CTO | Building Exceptional Technology

 Bratislava, Slovakia
sbatman.com
Joined September 2010
375 Photos and videos
Best bit about AI, I'm being productive even when I'm on the shitter 馃毥
5
1
14
489
When your collection of interlinked notes has grown so vast you can almost start building a galaxy. Adding useless but pretty visualizations to things is the best part of any project
80
Steven (Batman) Batchelor-Manning retweeted
Charles Curran
@charliebcurran
Sep 29
I used AI to explain the AI Doomer drama, with Lord of the Rings and kittens.
351
975
6,178
532,571
Looking at solutions to make your agents smarter? Getting the memory system right and getting the data to the agent at the right time is critical. There are a few ways of achieving this.
Steven (Batman) Batchelor-Manning
@S_BatMan
May 21
Article

Almost every serious memory system made the same retrieval decision. Here's why

Ask a flat-vector retrieval system to find the note that mentions the string `idx_memory_units_text_search`. The embedding model has no privileged representation for an arbitrary identifier. The

4
14
571
Different profiles allows you to quickly lean on the best model for the job across subscriptions / keys / local hardware. Automatic fallback, keeps the conversation going during outage or limits being reached
7
311
You only get so much value out of that 1M context, and its a lot less than you think
Steven (Batman) Batchelor-Manning
@S_BatMan
Jun 28
Article

You bought a 1M context. You got 50x less than you paid for.

Every vendor's headline context number is a lie. Not a small lie. A 50 to 100 times lie, depending on what the model is being asked to do. The architecture accepts the input. The model does not read

1
5
443
Are you using the right harness ?
Steven (Batman) Batchelor-Manning
@S_BatMan
Jul 4
Article

You don't have a context problem. You have a harness problem

The harness effect on real workloads runs 5% to 40%. The model effect between frontier models on the same benchmark runs 5% to 15%. The scaffold matters more than what's inside it. The cleanest single
1
139
This particular agent has supplemented and boosted our teams output drastically since it came online in may. delivering multiple features to the platform, managing ops and leading review etc. Scary how much has changed.
1
5
268
Steven (Batman) Batchelor-Manning retweeted
Matt Pocock
@mattpocockuk
Sep 23
Stop caring so much about model releases. Focus on the harness, and improving the environment your agent operates in. You'll find yourself far ahead of the curve.
142
147
3,962
168,354
I do like a clean ui for quick understanding of where my spend is
1
7
219
Ever feel like them limits are coming up quick ?
Steven (Batman) Batchelor-Manning
@S_BatMan
May 30
Article

You are not running out of tokens. You are wasting them. Here is the difference.

The longer ive been working with llm's the more that I find myself in the eternal battle between enough context, too much context, token wastage and compaction warfare, looking to agentic memory
6
374
Same Model different harness, Huge step up in capability. What's your go to Harness/IDE and why ?
Steven (Batman) Batchelor-Manning
@S_BatMan
Jul 4
Article

You don't have a context problem. You have a harness problem

The harness effect on real workloads runs 5% to 40%. The model effect between frontier models on the same benchmark runs 5% to 15%. The scaffold matters more than what's inside it. The cleanest single

2
4
424
Steven (Batman) Batchelor-Manning retweeted
Steven (Batman) Batchelor-Manning
@S_BatMan
Jul 1
Article

Long context isn't a capability. It's a standing charge.

One active user on Llama 3 70B at 128K context takes 40GB of GPU memory. That's half an H100. Per token, the cache footprint runs 320KB in BF16. Every additional user you put on that GPU costs you

2
6
68
433,139
AI may have replaced intelligence but it is yet to replace wisdom
chanduu
@chanduuu_cs
Sep 20
AI coding tools are making junior developers ridiculously fast. But there's a catch. If you don't understand the code, you can also become ridiculously fast at creating problems you don't know how to fix.
1
233
Bad take. This is not going to be the case anymore. If you're reviewing code manually line by line, you're wasting valuable human time. Build trust in your tools, have fail-safes and layers, and replace a people problem with an unlimited-clock AI problem.
kache
@yacineMTB
Sep 21
If you don't need to read every line of code that you are working on and just ship whatever the AI produces, you are working on something that doesn't really matter
1
187
Steven (Batman) Batchelor-Manning retweeted
Ankan Chowdhury
@AnkanXplorer
Sep 21
AI is getting smarter every month. humans still forget their passwords every week. who鈥檚 really evolving faster ?? 馃槄 AI or humans,
42
5
55
1,652
I love all the comparisons between AI and Gambling, and as someone who's in the know on both, i can assure you that they are to a point somewhat true, but not as much as some want you to believe.
1
1
172
One of the key strengths of having multiple agents on one problem even if they are using the same.model is you reduce down the impact of randomness and strengthen the overall output. Done correctly this saves you time and tokens.
1
3
195
Even with agents reading the llm/ai news and summarising it for me, I'm soon going to need a summary summary agent everything is moving so fast. The onlynsingularitynwe are going to reachbisnthe information one where we produce something "new" it all becomes impossible to consume
3
4
164
The trickiest part of building out any harness infrastructure is the continued battle of making some models more capable and accurate without making other models less capable, its a tricky art
1
135
Load more