Ghoshan (@ghostyfreak) header
Ghoshan (@ghostyfreak) profile picture

Leading @TaxxaAI for the battle of accounting AI

Colombo
Joined November 2024
Photo by Ghoshan (@ghostyfreak)Photo by Ghoshan (@ghostyfreak)Photo by Ghoshan (@ghostyfreak)Photo by Ghoshan (@ghostyfreak)Photo by Ghoshan (@ghostyfreak)Photo by Ghoshan (@ghostyfreak)Photo by Ghoshan (@ghostyfreak)Photo by Ghoshan (@ghostyfreak)Photo by Ghoshan (@ghostyfreak)Photo by Ghoshan (@ghostyfreak)
Oh boy, I'm not finding any reason to use gpt now. And thats really bad. Imagine @AnthropicAI won. no open model. no openai. Super intelligence in the hands of the very few.
1
15
Jensen might be the best real life rendition of Icarus
17
Mullaitivu grew faster than any district in Sri Lanka since 2012: +33%. In 2019 it also had the highest poverty rate: 44.5%. Asked in Claude, answered from census tables through our MCP.
19
Every MCP tool result should tell the model what to call next. ◇ `next`: ready-to-send calls ◇ `why` in plain words ◇ `matched_on` says when a synonym matched Fewer dead ends, fewer made-up field names.
18
Every data platform should ship an MCP before it ships a dashboard @developcountry
2
26
Type "power cuts". Get "Electricity supply reliability". ◇ One search parser for the site and the MCP ◇ Plurals stemmed, filler words dropped ◇ Synonyms always rank below typed words Nobody searches in government table titles.
2
39
This is how I measure what every MCP tool call costs. ◇ Proxy over D1, R2 and assets ◇ One metered env per call ◇ One Analytics Engine point per call
32
Attention is all you need applies to human more than agents
1
22
Everyone’s at the mercy of their barber
2
27
Deepseek flash 1017 is a really good model for our benchmark. Its way cheaper and cites accurately. Now only if someone could make a token efficient version of this model?
35
Idk what’s wrong with me. I prefer how open ai models talk. Miles better.
I hate talking to OpenAI models now. Opus 5.5 spoiled me.
37
Models keep absorbing the harness. Half your tool clarifications and “remember to…” rules are now dead weight. Ablate them: leave-one-out on a real eval set, delete what doesn’t move the score. Your agent outgrew its harness. Audit for redundant instructions Harness²’s probe-and-compare loop is a good pattern, but point it at deletions and grade against real labels.
2
37
@bot can you do this for for our harness
1
18
Ours too. Something is happening
here are our teams commits since january something is happening
16
The fall of @conductor_build needs to be studied. If they had just open sourced, bet on the community, ship at the pace of progress of the community.
1
34
Im going ahead with @bunjavascript check for the 2x speedup. Its the fastest. Already part of the toolchain. @voidzerodev's tsrs complains about circular type errors
1
96
Ghoshan retweeted
I think people will generally underestimate how much technological acceleration is directly attributable to deepseek. Because of them, it is going to proliferate. Really, it's because of them. Thank you, for everything
109
395
56
6,696
219,693