Mitesh B Ashar

30K posts

Mitesh B Ashar

Mitesh B Ashar

@iMBA

NOT an MBA.

Kolkata, India Katılım Ağustos 2007
2.7K Takip Edilen2.8K Takipçiler
Mitesh B Ashar retweetledi
antirez
antirez@antirez·
Markdown vs HTML. Every time we go from a semantically dense to a semantically sparse format, we lose. Even more today where less tokens from the same content is way better. I can understand we need a better markdown. I can't understand we should replace it with HTML.
English
176
89
1.5K
158.6K
Mitesh B Ashar retweetledi
Arnav Gupta
Arnav Gupta@championswimmer·
"When in doubt, let the LLM think" Usually a good policy for many things in the harness - esp. when deterministic answers are not available. I let the LLM decide when to prune its context, and it pruned it 5 times in a 96-turn task. Those turns cache hit drops to <70% but immediately recover from the next turn. Overall total cumulative cache hit of tokens in this entire task is still above 98% If this prune didn't take place this session would be 484k tokens (outside GPT 5.4 400k window and would have hit a compaction). But it actually ended the final turn at only 163k context window fill (only 40% of the window).
Arnav Gupta tweet mediaArnav Gupta tweet media
English
5
5
91
8.7K
Mitesh B Ashar
@championswimmer Ah! Just noticed you have the Opex quoted on the racks. Would help to have them quoted on the DCs too.
English
0
0
0
9
Mitesh B Ashar
@championswimmer SA East is a tricky one. Low costs, higher margins - super high taxes! Would take much effort to start doing so for the racks now! 🫣
English
1
0
0
14
Arnav Gupta
Arnav Gupta@championswimmer·
Made a CLI for Datacenter Tycoon, and making Claude play it. This can turn into vendingbench 🤣
Arnav Gupta tweet mediaArnav Gupta tweet media
English
4
1
68
6.5K
Mitesh B Ashar
@dexhorthy Humans:HTML + Agents:Markdown is the natural consensus. I like the middle ground of using Macdown 3000, when I want a slightly better view or export HTML/PDF to send to someone. That reserves the need to do an interactive HTML page only when I actually need that experience.
English
0
0
0
43
dex
dex@dexhorthy·
I think this is probably why a bunch of ppl in the comments are advocating for MDX - complex render logic lives outside the markdown, but can be displayed inline with mostly-token-efficient content
English
1
0
16
1.7K
dex
dex@dexhorthy·
Been helping users generate HtML to understand complex plans and changes for months now. But the real value of markdown is twofold, and html only hits the first one 1) clear, compact summary of a lot of information or intent - higher leverage faster easier understanding for HUMANS 2) a token efficient compact summary of information or intent for MODELS it’s important to know what your goal is - more leverage for humans, or more performance from models/agents If you need 2) then HTML is gonna blow up your context window way faster than 1) We often find ourselves using both, markdown for most cases, supplemental html if there’s an opportunity to make content more digestible for humans
Thariq@trq212

HTML is the new markdown. I've stopped writing markdown files for almost everything and switched to using Claude Code to generate HTML for me. This is why.

English
24
14
176
20.9K
Mitesh B Ashar retweetledi
Mitchell Hashimoto
Mitchell Hashimoto@mitchellh·
AI slop is good, actually. Slop is what enables fast parallel experimentation. The etiquette and skill is understanding the boundaries of where slop exists and the extent to which it should be cleaned up and how. A few examples: I’m working on the internals of some system right now. The API and GUI of this thing is fully zero shame slop. It’s horrible. But it lets me focus on the core quality while shipping a usable piece of alpha quality software to testers (transparent about the slop frontend). Similarly, this system has plugins. We sent agents in Ralph loops overnight to generate dozens of plugins. The plugins are slop. The quality is bad. The plugin API/SDK is absolutely not done. But we can test a full GUI with a full plugin ecosystem. When we change the API, we can regenerate them all. The cost of change is just tokens, the velocity is incomparable to before. I built Terraform. We tested and shipped TF 0.1 with about 3 very weak providers. Because we ran out of time. Building was slow. And when we changed our SDK the cost was immense. Totally different today, 10 years later. Today, I would’ve slop generated 100 providers (again, with transparency and cleanup later, but just to prove it out). As an anti example, I would not PR this (without prior warning) to another project. I would not throw this onto customers without full review or transparency (as I’m already doing). I would not accept first pass slop. It’s almost never right. Slop is a tool. And like anything else it’s not blanket bad or good. The context is everything.
English
107
218
2.6K
200.1K
Mitesh B Ashar
Mitesh B Ashar@iMBA·
Here is a demo of something I have currently been working on: Using Claude Code with @vercel's agent-browser for some browser automations in @ZohoBooks. I have bundled majority of the automation into a client side IIFE. This led to a 18x cost-saving as shown below, though realistically I feel it must be around 10-12x. A follow up to the demo demonstrates the advantage of being deterministic-first when building agentic systems. Post-demo: Claude code figured out how to batch the automation into a single run. Following that, I made the batching a part of my own automation.
Mitesh B Ashar tweet mediaMitesh B Ashar tweet mediaMitesh B Ashar tweet media
English
3
1
3
125
Mitesh B Ashar
Mitesh B Ashar@iMBA·
@kargarisaac @vercel @ZohoBooks I know that feeling. Avoiding it is such a naturally human response, given we have access to such a capable system now. OTOH, they're equally capable of giving us a better pov and clarity about the problem we want to solve and that helps me a lot in such a scenario.
English
1
0
1
25
Isaac Kargar
Isaac Kargar@kargarisaac·
@iMBA @vercel @ZohoBooks Makes sense. That's an important point. These days we avoid this because we don't know what we want and hpe the agent can do sth and figure this out 👌
English
1
0
0
27
Isaac Kargar
Isaac Kargar@kargarisaac·
@iMBA @vercel @ZohoBooks nice. deterministic-first is the right instinct. Let the agent discover and adapt, then graduate stable steps into code so future runs are cheaper and more reliable.
English
1
0
0
26
Mitesh B Ashar
Mitesh B Ashar@iMBA·
2. Even if (1) passes, is it large and multi step? Does it necessarily have to run in a single session? If its large enough, loading step by step will keep the agent much more grounded.
English
0
0
0
20
Mitesh B Ashar
Mitesh B Ashar@iMBA·
1. Will all of it be required anyway? Effectively, will my agent have to unconditionally read each of the files anyway, after I've broken them into multiple.
English
1
0
0
19
Mitesh B Ashar retweetledi
PunsterX
PunsterX@PunsterX·
BJP workers are so enthusiastic, they have started development work in West Bengal even before the govt has been formed.
English
181
1.7K
6.3K
189.7K
Mitesh B Ashar
Mitesh B Ashar@iMBA·
Remember Amazon streaming private conversations with Alexa a decade back? This has always gotten through. And unfortunately, keeps getting through.
Evan Luthra@EvanLuthra

🚨META’S SMART GLASSES ARE RECORDING YOU IN YOUR MOST INTIMATE MOMENTS.. AND SENDING ALL OF IT TO WORKERS IN KENYA WHO WATCH EVERY SECOND.. THEN META FIRED 1,108 OF THEM FOR TALKING ABOUT IT.. Swedish journalists discovered that footage from Meta’s Ray-Ban smart glasses is being sent to a facility in Nairobi, Kenya.. Where workers manually watch and label everything the glasses capture.. Not AI watching.. Humans watching.. Over 30 workers confirmed what they see every day.. People in intimate situations.. People on the toilet.. People undressing.. Credit card numbers.. Banking passwords.. Private messages on phone screens.. All completely visible.. One worker said.. “I don’t think they know, because if they knew they wouldn’t be recording”.. Meta marketed these glasses as “built for your privacy”.. “You’re in control of your data and content”.. The AI features cannot function without sending your footage to Meta’s servers.. There is no local option.. If you use the AI.. Your private life leaves your device.. Swedish journalists visited 10 retail stores.. Every single sales rep incorrectly told customers all data stays on the phone.. Not one knew the footage goes to Kenya.. Meta claims face-blurring protects identities.. Workers say it barely works.. Faces fully visible in low light, fast movement, complex backgrounds.. People in your bedroom.. Fully visible.. To strangers making $1.50 an hour.. Workers said the facility was “saturated with content that could trigger enormous scandals if leaked”.. So the company put them under constant camera surveillance and banned personal devices.. Workers surveilled to prevent them from leaking the surveillance footage they were watching.. Then the investigation went public.. Meta terminated the entire contract.. Claimed Sama “didn’t meet our standards”.. Sama fired back.. “At no point were we notified of any failure to meet those standards”.. 1,108 Kenyan workers.. Fired.. Six days notice.. Labor activists called it retaliation.. “The workers who trained the AI saw everything.. Owned nothing.. And lost their jobs the moment they spoke about it”.. 55% of these workers report clinical distress.. 52% meet thresholds for major depression.. They earn $1.50 an hour.. Meta made $56.3 billion last quarter.. The head of the Data Labelers Association said it best.. “It is African Intelligence powering European intelligence.. Which they are now calling Artificial Intelligence”.. Meta has sold 7 million of these glasses.. Targeting 10 million by year end.. A class-action lawsuit has been filed.. Kenya’s courts ruled Meta can be sued directly.. 200 former workers are pursuing a $1.6 billion claim.. 7 million cameras on 7 million faces.. Sending everything to the cheapest labor market they can find.. And they called it “built for your privacy.”

English
0
0
0
53
Mitesh B Ashar
Mitesh B Ashar@iMBA·
@nixxin While my 'sensed' reduction is just empirical, I can definitely say that 5.5 is certainly a large guzzler. Nevertheless, the delta is supposedly still quite large. x.com/bettercallsalv…
Thiago Salvador@bettercallsalva

@arankomatsuzaki That 21x usage ratio with Codex still hitting limits before Claude Max is worth noting for anyone picking plans. Codex's limits seem calibrated for a different baseline than heavy research use. Have you compared actual throughput caps or just the warning triggers?

English
0
0
0
7
Mitesh B Ashar
Mitesh B Ashar@iMBA·
@nixxin TBH, while codex is certainly more generous, I do 'sense' similar reduction trends on codex too.
English
1
0
0
7
Nikhil Pahwa
Nikhil Pahwa@nixxin·
working demo of 3rd desktop app is ready. built with opencode + Kimi K2.5 with 3x limits :D Is it just me or is Anthropic squeezing limits for sonnet 4.6? Hit a token limit in 40 min today (creating an LLM Wiki for 115 content files) for easy retrieval. If this continues, planning to move to DeepSeek V4 Pro. I'm still looking for a mac mini btw.
English
3
0
9
1.3K