@eito80.bsky.social
Merry Christmas from 🇯🇵🇯🇵🌲🎁
Grateful to all the support and kindness from people around this year 🙏
Grateful to all the support and kindness from people around this year 🙏
December 26, 2025 at 4:48 AM
Merry Christmas from 🇯🇵🇯🇵🌲🎁
Grateful to all the support and kindness from people around this year 🙏
Grateful to all the support and kindness from people around this year 🙏
@aquavoice_ wrapped 2025, kinda surprised texts, slack, cursor are so close within eachother in terms of words spoken
December 16, 2025 at 12:00 PM
@aquavoice_ wrapped 2025, kinda surprised texts, slack, cursor are so close within eachother in terms of words spoken
@cursor_ai has really cooked with this
I am churning off of lovable/replit
I am churning off of lovable/replit
December 11, 2025 at 4:04 PM
@cursor_ai has really cooked with this
I am churning off of lovable/replit
I am churning off of lovable/replit
Nano Banana 🍌 is truly amazing
November 25, 2025 at 12:00 PM
Nano Banana 🍌 is truly amazing
This is my favourite Japanese snack they sell in convenience stores (e.g. 7/11), LinkedIn Jelly
October 23, 2025 at 11:00 AM
This is my favourite Japanese snack they sell in convenience stores (e.g. 7/11), LinkedIn Jelly
Agents become dangerous when the "lethal trifecta" (h/t @simonw) conditions are fulfilled, where an attacker can hijack the agent to exfiltrate your private data!
OpenEdison tracks per-session agent tool use to block write actions when they encounter the lethal trifecta
OpenEdison tracks per-session agent tool use to block write actions when they encounter the lethal trifecta
October 1, 2025 at 4:11 PM
Agents become dangerous when the "lethal trifecta" (h/t @simonw) conditions are fulfilled, where an attacker can hijack the agent to exfiltrate your private data!
OpenEdison tracks per-session agent tool use to block write actions when they encounter the lethal trifecta
OpenEdison tracks per-session agent tool use to block write actions when they encounter the lethal trifecta
Introducing OpenEdison by @edison_watch, the AI Agent Firewall
Agents + Tools/MCP = Data leak risk
OpenEdison is an OSS firewall that deterministically blocks data exfiltration & dangerous agent action, even if jailbroken.
👇 comment your MCP use, I'll dm how risky your use is
x.com/Eito_Miyamu...
Agents + Tools/MCP = Data leak risk
OpenEdison is an OSS firewall that deterministically blocks data exfiltration & dangerous agent action, even if jailbroken.
👇 comment your MCP use, I'll dm how risky your use is
x.com/Eito_Miyamu...
October 1, 2025 at 4:11 PM
Introducing OpenEdison by @edison_watch, the AI Agent Firewall
Agents + Tools/MCP = Data leak risk
OpenEdison is an OSS firewall that deterministically blocks data exfiltration & dangerous agent action, even if jailbroken.
👇 comment your MCP use, I'll dm how risky your use is
x.com/Eito_Miyamu...
Agents + Tools/MCP = Data leak risk
OpenEdison is an OSS firewall that deterministically blocks data exfiltration & dangerous agent action, even if jailbroken.
👇 comment your MCP use, I'll dm how risky your use is
x.com/Eito_Miyamu...
You're absolutely right!
September 12, 2025 at 11:00 AM
You're absolutely right!
Madrid over the weekend
September 8, 2025 at 11:00 AM
Madrid over the weekend
Waymo footage from 2 weeks ago in SF
Cannot believe how well it drives
Cannot believe how well it drives
September 1, 2025 at 11:00 AM
Waymo footage from 2 weeks ago in SF
Cannot believe how well it drives
Cannot believe how well it drives
"The $1T Opportunity to Build the Next Amazon in Retail" by
@Konstantine is incredibly written
On-demand (e.g. Uber Eats) vs consultative shopping, MCPs for standardisation, Predictive Shipping, robot roaming stores, IoT-smart home proactive purchases, so much more
x.com/Konstantine...
@Konstantine is incredibly written
On-demand (e.g. Uber Eats) vs consultative shopping, MCPs for standardisation, Predictive Shipping, robot roaming stores, IoT-smart home proactive purchases, so much more
x.com/Konstantine...
August 19, 2025 at 11:00 AM
"The $1T Opportunity to Build the Next Amazon in Retail" by
@Konstantine is incredibly written
On-demand (e.g. Uber Eats) vs consultative shopping, MCPs for standardisation, Predictive Shipping, robot roaming stores, IoT-smart home proactive purchases, so much more
x.com/Konstantine...
@Konstantine is incredibly written
On-demand (e.g. Uber Eats) vs consultative shopping, MCPs for standardisation, Predictive Shipping, robot roaming stores, IoT-smart home proactive purchases, so much more
x.com/Konstantine...
this is a hilarious slide
August 12, 2025 at 11:00 AM
this is a hilarious slide
If you don't have this for your bed, can you really call yourself a finance bro?
August 6, 2025 at 11:00 AM
If you don't have this for your bed, can you really call yourself a finance bro?
OAI IMO Interview insights:
- The OAI model that solved the IMO can tell if it "doesn't know" how to solve the problems, like it did on Q6 -> huge???? Combined w/ Anthropic steering research, hallucinations might be solvable???
- Emphasis on progress on non-verifiable tasks
- The OAI model that solved the IMO can tell if it "doesn't know" how to solve the problems, like it did on Q6 -> huge???? Combined w/ Anthropic steering research, hallucinations might be solvable???
- Emphasis on progress on non-verifiable tasks
August 5, 2025 at 11:00 AM
OAI IMO Interview insights:
- The OAI model that solved the IMO can tell if it "doesn't know" how to solve the problems, like it did on Q6 -> huge???? Combined w/ Anthropic steering research, hallucinations might be solvable???
- Emphasis on progress on non-verifiable tasks
- The OAI model that solved the IMO can tell if it "doesn't know" how to solve the problems, like it did on Q6 -> huge???? Combined w/ Anthropic steering research, hallucinations might be solvable???
- Emphasis on progress on non-verifiable tasks
When flocks of Technology Brothers in SF flood Anthropic with a bunch of requests through Claude Code ,and you can't even ask basic questions anymore ☠️☠️☠️
July 22, 2025 at 11:00 AM
When flocks of Technology Brothers in SF flood Anthropic with a bunch of requests through Claude Code ,and you can't even ask basic questions anymore ☠️☠️☠️
Couldn't believe it when I saw this
Google is losing search market share so badly they have to start putting ads in London underground
First time in my life I saw a Google ad
Google is losing search market share so badly they have to start putting ads in London underground
First time in my life I saw a Google ad
July 21, 2025 at 4:00 PM
Couldn't believe it when I saw this
Google is losing search market share so badly they have to start putting ads in London underground
First time in my life I saw a Google ad
Google is losing search market share so badly they have to start putting ads in London underground
First time in my life I saw a Google ad
Incredible blog post
July 14, 2025 at 11:00 AM
Incredible blog post
I think my most heretical idea that I think is true is that 80% of human to human communication can be autocompleted by an AI
The meme turing test has already been solved too. What's the reason to believe it cannot be solved for human-to-human communciation?
The meme turing test has already been solved too. What's the reason to believe it cannot be solved for human-to-human communciation?
July 11, 2025 at 11:00 AM
I think my most heretical idea that I think is true is that 80% of human to human communication can be autocompleted by an AI
The meme turing test has already been solved too. What's the reason to believe it cannot be solved for human-to-human communciation?
The meme turing test has already been solved too. What's the reason to believe it cannot be solved for human-to-human communciation?
great @latentspacepod ep with @polynoamial the GOAT
Insights:
- Deep Research is an example of non verifiable RL & many apps exist where RL + application layer is useful
- "existing multi-agent RL is not bitter lesson pilled enough", comments on not needing explicit external agent modelling
Insights:
- Deep Research is an example of non verifiable RL & many apps exist where RL + application layer is useful
- "existing multi-agent RL is not bitter lesson pilled enough", comments on not needing explicit external agent modelling
July 10, 2025 at 1:35 PM
great @latentspacepod ep with @polynoamial the GOAT
Insights:
- Deep Research is an example of non verifiable RL & many apps exist where RL + application layer is useful
- "existing multi-agent RL is not bitter lesson pilled enough", comments on not needing explicit external agent modelling
Insights:
- Deep Research is an example of non verifiable RL & many apps exist where RL + application layer is useful
- "existing multi-agent RL is not bitter lesson pilled enough", comments on not needing explicit external agent modelling
Recent YC video is now talking about meta-prompting & evals, but they still don't know about @DSPyOSS @lateinteraction
YC clearly 3 steps behind
YC clearly 3 steps behind
June 7, 2025 at 6:10 PM
Recent YC video is now talking about meta-prompting & evals, but they still don't know about @DSPyOSS @lateinteraction
YC clearly 3 steps behind
YC clearly 3 steps behind
Why vent to a bartender in a night bar when you can vent to a robot bartender in Prague (at 4pm)
June 1, 2025 at 10:53 PM
Why vent to a bartender in a night bar when you can vent to a robot bartender in Prague (at 4pm)
Visiting @alexocheema & @mo_baioumy at @exolabs office
I swear they're going to get kicked out of the WeWork for suspicions of abusing free electricity for bitcoin mining or causing a fire with the radiator-levels of heat coming from these mac minis
I swear they're going to get kicked out of the WeWork for suspicions of abusing free electricity for bitcoin mining or causing a fire with the radiator-levels of heat coming from these mac minis
May 21, 2025 at 4:00 PM
Visiting @alexocheema & @mo_baioumy at @exolabs office
I swear they're going to get kicked out of the WeWork for suspicions of abusing free electricity for bitcoin mining or causing a fire with the radiator-levels of heat coming from these mac minis
I swear they're going to get kicked out of the WeWork for suspicions of abusing free electricity for bitcoin mining or causing a fire with the radiator-levels of heat coming from these mac minis
2 extremely underrated models not getting enough noise
1) Gemini-2.0-Flash >>> gpt-4o-mini or any other models
2) (Groq) Llama-3.3-70B / DeepSeek-R1-Llama-70B distill (smarter, faster, cheaper than gpt-4o, though caveat that Groq is VC-subsidised)
source: @ArtificialAnlys
1) Gemini-2.0-Flash >>> gpt-4o-mini or any other models
2) (Groq) Llama-3.3-70B / DeepSeek-R1-Llama-70B distill (smarter, faster, cheaper than gpt-4o, though caveat that Groq is VC-subsidised)
source: @ArtificialAnlys
March 11, 2025 at 12:00 PM
2 extremely underrated models not getting enough noise
1) Gemini-2.0-Flash >>> gpt-4o-mini or any other models
2) (Groq) Llama-3.3-70B / DeepSeek-R1-Llama-70B distill (smarter, faster, cheaper than gpt-4o, though caveat that Groq is VC-subsidised)
source: @ArtificialAnlys
1) Gemini-2.0-Flash >>> gpt-4o-mini or any other models
2) (Groq) Llama-3.3-70B / DeepSeek-R1-Llama-70B distill (smarter, faster, cheaper than gpt-4o, though caveat that Groq is VC-subsidised)
source: @ArtificialAnlys