
AI news, tools, and breakthroughs. MSDAE Graduate researcher @Northeastern
Eastern Seaboard
Joined March 2011
- Tweets7.6K
- Following498
- Followers1.3K
- Likes7.1K
Pinned Tweet
Charlie Wells with @Bloomberg reached out to me to ask about my thoughts on VCX and publicly traded funds that hold private AI and tech company shares.
After confirming he was not a Russian bot (always important financial due diligence) I had a great conversation about a trend I think more people will start paying attention to:
Public-market investors are looking for ways to gain exposure to private AI and tech companies before they IPO.
My interest in AI started years ago after reading The Singularity Is Near, which completely changed how I thought about technology and society.
The next five years in AI are going to be wild.
Link to the Bloomberg article below
I create a ton of cool stuff in the ChatGPT Codex app, only to lose my train of thought a day or week later. There has to be a better way to pick up where I left off than the current UI allows.
Right now the best way to get myself back on track is describing it to my dot until it understands and takes me back.
61
Gemini 4 Neon scores 61 on the Artificial Analysis Intelligence Index at xhigh, third overall behind GPT-6.2 Astra and Claude Fable 5.2, both at max. Price: $3 / $15 per million tokens.
○ Cost per task: $6.44 vs $7.76 for Gemini 4 Carbon at list price; about 21 minutes vs 36
○ AA Coding Agent Index in Antigravity CLI: 59, vs 58 for Claude Code with Fable 5.2 or Opus 5.5
○ Coding agent cost per task: $6.86 vs $12.61 for Opus 5.5 and $15.47 for Fable 5.2
○ Cached input $0.15, a 95% discount; 1M token context; four thinking levels
○ Paid API tier only; also in Antigravity and the Gemini app for Pro and Ultra
Google's own figures: Vals Index 70.9% (#2), Zapier AutomationBench 56.3% (#1).
I wonder if Sam Bankman-Fried is going to celebrate the $ANT IPO in Prison today
153
Microsoft shipped seven MAI models at Ignite on Nov 17. The pitch is cost per task, not leaderboard rank.
1) MAI Code 2: 98% of GPT-6.1 Sol's score at about 40% of the price, averaged over DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1
2) Same average: 96% of Claude Opus 5.5 at 5% of the price
3) GitHub Copilot rates: Code 2 $0.50 / $2.50, Code 2 Flash $0.05 / $0.25 per 1M tokens, promotional through May 17, 2027
4) MAI Cyber 2 Flash beats GPT-6 Astra on CyberGymE2E at under 2% of the cost per task
5) MAI Realtime 1 speech to speech starts at $0.01 per audio minute
Microsoft says Code 2 is not yet tuned for science research or very long horizon agent work. Every comparison above is vendor reported.
158
GPT-6.2 Astra lists at $8 input and $40 output per million tokens, 20% below GPT-6 Astra.
• 1,050,000 token context, 922,000 max input, 128,000 max output.
• Artificial Analysis Intelligence Index: Astra 62 (second behind Claude Fable 5.2 at 65), Sol 57.
• Independent cost per Intelligence Index task: Astra $4.00 versus $6.24 for GPT-6 Astra (36% lower).
• Available in ChatGPT Work, Codex, and API as gpt-6.2-astra and gpt-6.2-sol for Plus, Pro, Enterprise, Edu.
• OpenAI rates both Critical for cybersecurity under its Preparedness Framework.
Independent scores show knowledge-work gains and flat or lower coding agent results relative to prior Astra and Sol tiers.
1
3
158
$NVDA This has to be one of the most anticipated earnings calls ever for any company. The former CEO posted that he would like to return in the morning, then the WSJ reported an SEC investigation at 3:57. The entire AI industry valuations are tied to the guidance in the call at 5PM EST. 🤞🏼
Between the story below and the top 10 banks all having problems simultaneously today fair to conclude a certain mid east country is upset with the US
94
Will be a crying shame if some variation of Johnny Flynn/Frankie Cooper is not incorporated as the next bond villain with the new direction MGM is going with Robert Pattinson as Bond
We can debate the impact of Jensen Huang’s epic, quick downfall and how it will affect Nvidia shareholders and the companies they invested in. But the devastation to the black leather jacket community may be unrecoverable.
1
2
113
Google DeepMind's AI agents generated about 6.5 million lines of machine checked Lean proof showing Signal's compiled code meets its core security properties. The verification took four weeks.
1. Target: libsignal v0.105.4, unmodified Rust compiled with stock rustc to AArch64 machine code
2. Coverage: the post quantum PQXDH handshake and Triple Ratchet; forward secrecy, authentication, post compromise security, handshake secrecy
3. Trusted base: about 26,000 lines of human reviewed specifications and AArch64 semantics; Lean's kernel checks the rest
4. Human review: the agents flagged about 1% of the work for a second look
5. Control test: two real, previously fixed bugs were reinserted; both times the system proved the buggy build insecure
Limits: cryptographic primitives are assumed secure, and scope is one to one sessions, not all of Signal. The Lean development is public on GitHub.
Great example of a stream-of-consciousness thought and the genesis of what today's tools can turn it into.
My AI agents found an unmapped Maya site in Mexico, using LiDAR data from a NASA dataset.
72 hours from a random thought on vacation to a reply from Mexico's national archaeology institute.
What a time to be alive…
46
OpenAI updated dots on Nov 9 so one dot can connect to multiple local computers, and you choose which machine handles a task.
▪ Each thread uses one computer at a time; start a separate thread on each machine for independent work
▪ Connect from the ChatGPT desktop app: open the dot profile, find that computer under Computers, then select Allow and Allow access
▪ The machine must stay online with the app open; Offline means unavailable, not revoked
▪ The cloud computer stays connected, and access applies from any channel, including a phone
▪ Local access is separate from Codex or Work Sync; neither of those connections alone grants it
A task does not split across two machines. Parallel work needs separate threads.
2
3
204
ByteDance Seed released Seedream 5.5 on Nov 9, an image model built around repeated edits on the same file.
1) Keeps unedited regions intact at the pixel level across up to 100 rounds of editing
2) Native RGBA output with full and partial transparency, no separate background removal
3) Expert GSB preference vs the prior version was positive in all four comparisons, up to +55.62% on image-to-image realism
4) ByteDance says it leads GPT Image 2.5 Sunburst and Nano Banana 2.1 on text structure in both text-to-image and image-to-image
5) A Pro version of the 5.5 series is announced, not released
Every comparison is ByteDance's own expert rating, and OpenAI shipped Images 2.6 on Nov 4, a newer release than the Sunburst version tested.
1
3
233
By the time you prepare yourself it’s already too late
Imagine you are put in charge of monitoring a massive water tank the size of a football stadium. The water inside doubles in volume every sixty seconds. At 12:00 noon, the tank overflows. When should you have become concerned? At 11:59 the tank was only half full.
1
3
340

















