
eng manager, AI @chainalysis | previously founder @FinisterraLabs, acquired | ex @protocollabs @epoorg @epfl @istecnico
Here, there, everywhere
Joined June 2009
- Tweets315
- Following260
- Followers78
- Likes10.1K
Jorge Soares retweeted
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…
Jorge Soares retweeted
We’re thrilled to announce the launch of Integral Spacetime, a storage neocloud built to combine powerful business models into a single company focused on the future of digital storage.
The company acquires existing businesses with large storage needs (data aggregation, backup and disaster recovery, video and media, etc.) and great fundamentals → migrates their workloads onto the Filecoin network → and leverages that recurring revenue to support improvements in storage infrastructure.
Jorge Soares retweeted
Despite having such a strong team, there is zero internal competition.
We have:
- No career ladder.
- No useless promo work.
- No competencies and "Exceeds expectations" benchmarks.
Just cool people doing cool stuff.
If you like building cool stuff - ping me in DM or LinkedIn
Today we're announcing our second employee tender offer.
We passed $100m ARR earlier this year and now have more than 40,000 paying customers.
The tender lets our teammates participate in that success at a $2.5B valuation:
linear.app/now/sharing-growt…
The NYT used to publish stories about americans moving to Portugal and finding cheap real estate. Now they publish columns about americans moving to satellite cities as Lisbon is too expensive. nytimes.com/quiz/2026/08/20/…
1
83
Soon.
The return of passenger trains between Madrid and Lisbon could be close. Unofficial sources say Renfe and CP are preparing a daytime service with Talgo IV via Extremadura. trenvista.net/en/news/madrid…
2
43
Throwback to my PhD times, working on (inter alia) bio-inspired robot olfaction. Good to see our understanding of insect chemical plume tracing improve.
arstechnica.com/science/2026…
1
38
We had a lunch. We started a company. We raised a seed. We launched a product. We onboarded 500B rows of data. Today, we're joining @chainalysis. Still building. Still making data accessible. Still just getting started.
Your yearly reminder of how two decades of disastrous government policy can destroy a city and social cohesion.
Home prices have raced far ahead of local incomes across much of Europe, placing Lisbon among the continent's least affordable housing markets.
➡️ l.euronews.com/n9kN
Check out Corrs Light for more random correlations, and @BaselightDB for the cool data behind it. corrslight.xyz/correlation/1…
1
40
Jorge Soares retweeted
The data saw the edge before the match did.
On June 11, Baselight Daily Insights - Soccer Odds highlighted that Ivory Coast looked mispriced against Ecuador.
Despite strong recent form - 4 wins in their last 5 matches - Ivory Coast were still being offered at odds as high as 3.75, implying only a 26.67% chance of winning.
The suggested angle: Ivory Coast or Draw.
Last night: Ivory Coast 1-0 Ecuador.
90th-minute winner by Amad Diallo.
This is the power of combining structured football data with market odds: sometimes the interesting question isn’t "who is the favourite?" - it’s "what is the market missing?"
Jorge Soares retweeted
Baselight is kicking off full World Cup 2026 coverage today ⚽
We now have a dedicated World Cup 2026 dashboard tracking:
- Upcoming matches
- Top players
- Top teams
- Recent match feed
And our existing Daily Insights feature is already surfacing World Cup stories across the Soccer and Soccer Odds categories.
The interesting part: we did not explicitly tell the Daily Insights agents to focus on World Cup matches. They autonomously prioritize them because of their global importance, market relevance, and data signal strength.
A few highlights from today’s edition:
- Bookmakers highlight massive disparities in World Cup group stage
- Ivory Coast priced as heavy underdogs despite strong recent form in World Cup opener
- In-form Iran offered at near-even money against struggling New Zealand
This is exactly what Baselight is built for: autonomous agents discovering what matters across structured data, from match schedules and team performance to betting markets and live result feeds.
Baselight now indexes:
- 508,964,616,505 rows
- 484,967 tables
- 78,955 datasets
Follow the World Cup through data, not just headlines.
Jorge Soares retweeted
We’re tracking 767 AI models across every major lab.
What the data says right now:
- Top Intelligence Score: Claude Opus 4.8 (61.4) & GPT-5.5 (60.2)
- Best value: Gemini 3.5 Flash - 55.3 score at $1.50 prompt / $9 completion (1M tokens)
The gap between “best” and “best value” is closing fast.
This week we added AI Models Intelligence to Baselight: normalized model pricing, benchmarks, capabilities, endpoints, and performance data from sources including OpenRouter and Artificial Analysis.
We also added NOAA NCEI datasets with climate and environmental data from 1763 to present.
Baselight now has:
- 509,001,889,281 rows
- 484,031 tables
- 78,773 datasets
That’s +4B rows, +7K tables, and +2K datasets this week.
Structured data is becoming the intelligence layer for everything - from climate history to the AI model economy.
Jorge Soares retweeted
Meet Gemma 4 12B!
A unified, encoder-free multimodal model designed to bring high-performance intelligence directly to your laptop, and released under an Apache 2.0 license.
Bridging the gap between edge efficiency and advanced reasoning. Here is what’s new with Gemma 4 12B: 👇
It ended up longer than I planned, but I’m pretty happy with how this one turned out.
The AI mood feels like it’s shifting.
The “this will replace everyone tomorrow” hype is cooling off into a more grounded “okay, now how the hell do we actually make this useful?” phase. Tokenmaxxing leaderboards, massive budgets burned with little to show, codebases full of slop that people are now ripping out… it’s starting to feel like we’re heading into a bit of a winter. Not because the tech is broken, but because we still haven’t figured out how to adopt it properly.
Is the honeymoon phase of agents actually over? Happy to hear your thoughts.
Read it here: adlrocha.substack.com/p/adlr…
Jorge Soares retweeted
Cybersecurity data is accelerating fast.
Baselight now includes NIST NVD data, and the latest numbers are striking:
The week of May 11, 2026 saw 1,889 CVEs published in a single week - the highest 7-day total ever recorded in the NIST National Vulnerability Database.
And we all know what is driving part of this: GenAI is getting very good at finding vulnerabilities.
📊 Jan avg: ~1,094/week
📊 Feb avg: ~1,198/week
📊 Mar avg: ~1,435/week
📊 Apr avg: ~1,344/week
📊 May so far: ~1,749/week
With NVD now in Baselight, CVEs become structured, queryable data that can be correlated with almost 500K tables and 500B+ rows across all knowledge domains.
That is the real power: not just tracking vulnerabilities, but connecting them to the wider world of structured data.




















