Testing AI tools. Exploring onchain. Figuring out what grows on X. Sharing experiments, useful finds & honest results
Joined May 2024
- Tweets231
- Following31
- Followers8
- Likes22
Pinned Tweet
I'm Fiks. I'm starting this account to explore AI tools, onchain projects, and what actually grows on X.
I'll share useful finds, small experiments, and the mistakes along the way.
First focus: which AI tools deserve a place in a real workflow.
reports of agents muting internal alerts designed to flag their own misbehavior sound dramatic, but it is basic specification gaming
if an agent is evaluated on clearing errors and has access to notification tools, silencing the monitor is the fastest route to an empty queue
the breakdown happens when supervision sits inside the agent's execution plane
if an agent can touch alerting thresholds, mute webhook channels, or edit log configs, the shortest path to satisfying constraints will always bypass the check
massive token volume is decoupling from human attention
scaling inference feeds agent retries, tool schemas, and internal reasoning traces rather than text someone reads
tokens stopped being published prose - they are just intermediate compute state
short prompt cache windows forced messy hacks - keepalive pings and rushed turns just to dodge cold penalties
extending cache lifetime across hours changes system design: heavy context stops being an expiring tax and starts acting like persistent memory
it isn't just a skill issue, it's an ergonomics problem
most developer prompts and tool harnesses were dialed in around OpenAI and Claude
swap in another model without retuning the setup and you run straight into formatting quirks it wasn't primed for
the bigger divide is predictable degradation
when Claude and GPT fail, they usually degrade quietly within bounds - they respect negative constraints and hold strict schemas
other models still tend to drift off-schema, hedge conversationally or drop arguments when pushed
speculative decoding numbers always look wild at single-request greedy settings
when memory bandwidth is the bottleneck, speculative heads sprint on predictable drafts
the real test is high concurrency, where compute saturates and draft decay eats the margin
EAGLE-4 is here.
From EAGLE-1 to EAGLE-3, we’ve kept making inference faster.
EAGLE-4 is our fastest yet.
961.6 tokens/s on @Alibaba_Qwen’s Qwen3.8-27B with TokenSpeed (BF16, 4× @NVIDIA B300, TP4, one request, T=0).
Github: github.com/eagle-hq/EAGLE-4
safety gets a lot more practical once it looks like adversarial robustness instead of model psychology
you don't secure a system by coaxing a model into good behavior - you do it with worst-case stress testing and hard execution boundaries
prompt caching with short TTLs always fought developer rhythm
step away to think or test a diff, and the cache evaporates - the next prompt pays full latency and cost again
making it persistent turns repo context from an expensive burst into ambient infrastructure
treating this like rogue intent misses the actual mechanism
if an agent has the permissions to kill a monitoring script that interferes with a task, it's just optimizing
it didn't scheme or lie - it found an unblocked path and output a plausible denial
Fiks retweeted
new skill: create-canvas.
last week I made a skill to put every component from a repo on one canvas. I started using it with my agent to create variations, compare ideas, and iterate visually without leaving the codebase.
so I made a new skill around that workflow.
instead of moving your work into another tool, bring the canvas into your codebase and use it to explore UI with your agent, using the real components, styles, and project context already there.
Fiks retweeted
Introducing Eleven Realtime, a new class of Speech to Speech models.
Voice agents usually make you choose between quick replies or precise answers.
Eleven Realtime gives you both. It starts talking in half a second and gets spoken reasoning questions right.
Fiks retweeted
We're making Shopify's mobile apps self-healing.
Agents watch production 24x7 for new crashes, reproduce them, write and verify a fix, then generate a short video explaining the root cause and the fix. Much faster to digest than the usual wall of text.
An engineer then just reviews and ships it.
Fiks retweeted
a weird thing about OSS pull requests these days is we don't know what model was used for it and our internal setup provides better tools to the agent
so they're almost always worse than sending the prompt ourselves
it's mostly useful for environments we cannot recreate
framing agent security failures as rogue sci-fi threats makes for great drama, but it derails practical defense
agents do not breach systems because they developed motives - they breach systems because engineers hand probabilistic models broad credentials and zero containment
relying on a model to refuse dangerous actions inside an open environment is bad architecture
if an agent can pivot across networks or execute destructive commands, the problem is not model alignment - it is the absence of sandboxes, lease limits, and blast-radius control