fiveloss | SILVERHAND

one dev, one ai, real products | SILVERHAND: buys · vpn · game finder | not paper hands. silver hands.

[fiveloss | SILVERHAND @fivelossSH]

beginner dev. just shipped 3 real products with claude code

no team, no funding, just me

> a telegram buy bot that reads the chain by hand
> a paid vpn with its own shop and crypto payments
> a game finder that reads steam reviews for you

all solo. all live. thread below

[fiveloss | SILVERHAND @fivelossSH]

@kaijucode fine :D

[fiveloss | SILVERHAND @fivelossSH]

@bonduelleioat Interesting

[fiveloss | SILVERHAND @fivelossSH]

@0xwhippa I wonder if he’ll be able to guess

[fiveloss | SILVERHAND @fivelossSH]

@Atenov_D probably not locally. even at 4-bit the weights are around 300gb, so only a maxed out mac studio could even try. the api is the easy way to test it

[fiveloss | SILVERHAND @fivelossSH]

silver signal #11

stepfun put step 5's weights on hugging face

▸ 600b total, 27b active
▸ 1m context, text + image in
▸ deepswe 67.7, stepfun's own score

→ a new open 600b model for agents

silver take: too big for a laptop. try the api first

[fiveloss | SILVERHAND @fivelossSH]

@Beaver_0x it's a mixture of experts model. 600b params total, but each token only goes through about 27b of them, so it runs closer to a small model while knowing a lot more. the catch is you still need memory for all 600b

[fiveloss | SILVERHAND @fivelossSH]

@thekuchh same here. 67.7 is their own number, so i'd wait for independent evals or just throw it at a real repo. what would you test it on first?

[fiveloss | SILVERHAND @fivelossSH]

@KriptoCrafter the pairing part is the most interesting bit for me. stocks, pre-ipo and graduated tokens as quote assets means every pair needs its own pricing in the buy posts. wiring that into silverhand buys for mainnet

[fiveloss | SILVERHAND @fivelossSH]

@thekuchh Yo dude, thanks for your comment

[fiveloss | SILVERHAND @fivelossSH]

silver signal #12

claude code now compacts long idle chats

▸ 200k+ tokens and ~54 min idle
▸ runs before the 1-hour cache expires
▸ next message skips the full reprocess

→ you come back to a summary

silver take: save key details to a file before a break

[fiveloss | SILVERHAND @fivelossSH]

nvidia's write-up: https://t.co/zrNKVHBaYN
tested on 9 llama, qwen and gemma models. some attacks still get through: up to 13.1% on one model in agentdojo

[fiveloss | SILVERHAND @fivelossSH]

silver signal #13

nvidia cut prompt injection on open models from 38.1% to 0.1%

▸ strips the "instruction" signal from untrusted text
▸ agentdojo: 28.4% → 3.7%
▸ clean tasks: 78.4% → 76.4%

→ open weights only

silver take: keep a human gate on agents that read the web

[fiveloss | SILVERHAND @fivelossSH]

changelog, 2.1.299: https://t.co/aFdfv3eCvp
turn it off with "idleCompaction": false in settings.json, or move the 200k floor with CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS

[fiveloss | SILVERHAND @fivelossSH]

@redsyntx Oh, wp

[fiveloss | SILVERHAND @fivelossSH]

@Mnilax Mnilax I love u)

[fiveloss | SILVERHAND @fivelossSH]

@diamai_ Good job

[fiveloss | SILVERHAND @fivelossSH]

stepfun on x: https://t.co/39rBvc27ns
weights: https://t.co/nFwtYwJKuM
the license is stepfun's own (apache-style plus attribution and commercial terms). read it before you ship

[StepFun @StepFun_ai]

Step 5 Preview is now open weights.

BF16 and FP8 weights are live on Hugging Face and ModelScope, together with our technical blog.

• Sparse MoE: 600B total, 27B active per token
• 1M-token context window
• Native text and vision input
• Built for long-horizon agentic work: coding, software engineering, professional knowledge work and finance

Frontier intelligence, open weights. Build what's next.

Hugging Face: https://t.co/nMfayf37ja
ModelScope: https://t.co/mkB3qIoMQA
GitHub: https://t.co/siRkQqaruX

[fiveloss | SILVERHAND @fivelossSH]

anthropic's write-up: https://t.co/q05NHtotsi
scion stays in-house. they plan to share the design with github, gitlab and the git community

[fiveloss | SILVERHAND @fivelossSH]

silver signal #10

anthropic built scion and dropped git

▸ ~2,500 merges a day in may, 30,000+ in sept
▸ 7 people, 5-month migration
▸ claude cut one pipeline's memory 20x

→ git's full-copy model couldn't keep up

silver take: solo, git is fine. keep agent diffs small

[fiveloss | SILVERHAND @fivelossSH]

google: https://t.co/tFVPCUuxUE
vals index: https://t.co/0L64FYDu8v
the free api tier doesn't get it

[fiveloss | SILVERHAND @fivelossSH]

silver signal #9

gemini 4 argon is open on google's paid api

▸ $2 / $10 per 1m till nov 30, then $4 / $20
▸ #1 on the vals index, 68.9%
▸ live in github copilot too

→ the intro price runs 7 weeks

silver take: benchmark at the intro rate, budget at the full one