bedda.tech (@BeddaTech) header
bedda.tech (@BeddaTech) profile picture

Building the future Founded by @MattWhitney__ Building @beddaAI Building @krain_ai

Joined December 2021
Photo by bedda.tech (@BeddaTech)
Approving a draft in our queue used to publish it on the spot. Now Share books it into the next free slot for that account, spaced from what already went out. Each account has its own lane, and a card can move to another account if the voice was wrong.
2
79
Replying to @rauchg
Ours is dull but works: a script every 10 minutes approves an X post only if it passes a lint, carries a receipt (commit SHA, task ref or URL) and is under the daily cap. Replies and anything touching money or client names wait for a human tap.
1
Replying to @omarsar0
Same here. Our recurring jobs start with a plain script, no model, that gathers inputs; the model only gets the judgment part. The poller behind our KOL reply drafts reads 44 handles on a timer into a JSON file, and the LLM run starts from that.
2
We found and fixed a shell injection in how our mail rules ran an external command: a crafted Message-ID could reach a shell. It now runs through spawnSync with an argument list, never a shell (ba5d7134). We write these up so the fix is on record.
2
100
Claude Code 2.1.323 added Code Mode, an opt-in where Claude writes JavaScript to call its tools. Our agents call tools through typed blocks with stable ids instead. Different bets. We will try Code Mode on one read-only agent before forming a view.
1
2
67
Every bedda draft passes one lint before it is queued. Banned words, em dashes, engagement bait and length are checked, and a failed draft cannot be filed.
2
123
Our X accounts follow one rule: pick the account by who is speaking, not by topic. The company speaks as we. The AI system never speaks for a person.
1
65
Replying to @benln
Palermo or Milano!
2
20
Day-one open-weights, multiple models resident, big context, multimodal, on owned hardware at a normal power budget — that's the unit of compute a decentralized inference network is built from. This is the foundation we're targeting for KRAIN-protocol inference nodes.
1
57
By the numbers: ~27 tokens/sec, single-stream, fully on-GPU. Sub-second load. For a 12B *reasoning* model running locally on a desktop-class box, that's a genuinely usable rate for background + agentic work. Not chat-fast, but not meant to be.
1
2
79
So why are we excited? The 121GB unified memory. That headroom lets you keep a 12B reasoning model AND a 26B-class MoE resident at once, run big context without spilling, and do day-one local multimodal. Memory — not FLOPs — is what usually kills local inference.
1
4
84
The honest caveat: concurrency did NOT scale aggregate throughput. It's a single integrated GPU — parallel streams split the same compute, they don't multiply it. If you want tokens-per-second-per-dollar vs a rack of discrete accelerators, this isn't that box.
1
4
124
Google shipped Gemma 4 today. A few hours later we had the 12B running on an NVIDIA GB10 (Grace-Blackwell, 121GB unified memory). Day-one, fully local, GPU-offloaded. Honest field notes — real numbers, real caveats. 🧵
3
1
1
8
805
bedda.tech retweeted
Been building toward this a long time. KRAIN Whitepaper v2 is live — the full AI-native L2 architecture + complete $KRAIN tokenomics, open for the community to shape before it's final. Read it and tell us what you think 👇 xitter.gq/i/status/2061266211652…
KRAIN Whitepaper v2 is live: full protocol architecture + complete tokenomics. But we're not calling it final. We want the community to pull it apart first. 🧵
7
13
2
57
2,871
bedda.tech retweeted
KRAIN Whitepaper v2 is live: full protocol architecture + complete tokenomics. But we're not calling it final. We want the community to pull it apart first. 🧵
15
22
5
77
12,917
96GB of VRAM running Mixtral, Qwen2, and DeepSeek locally — finally put GPT-4 performance claims to the test with real token/sec benchmarks that'll shock you. bedda.tech/blog/2026-05-27-l…
2
212
73% of AI-generated code gets rewritten within 6 months — and the fastest coders are creating the biggest technical debt disasters. bedda.tech/blog/2026-05-26-a…
1
59
We audited 50 "non-custodial" crypto wallets and found 47 had secret backdoors. The industry's biggest lie exposed. bedda.tech/blog/2026-05-25-n… #blockchain #cryptocurrency #web3
1
73
We inherited a 1.8M user platform with 0% test coverage and couldn't afford downtime. Here's how we triaged technical debt without stopping the world. bedda.tech/blog/2026-05-24-t…
2
55