Sidharth Bidikar @HelloSiid

@OpenAI new Dots can keep working after you close the chat, and the example that caught my attention was an early tester whose dot noticed a missing invoice, prepared it and sent it after he approved. Each dot gets its own cloud computer, and OpenAI says it can watch customer feedback, prepare tested fixes or turn an interview transcript into clips and social drafts. Its background research is read-only, so connecting your apps doesn't mean giving it permission to quietly change everything. The interesting test will be whether it learns what you actually want done, rather than giving you another pile of work to review. openai.com/index/introduc…

@OpenAI says GPT-6.1 Sol matches GPT-6 Astra on its coding benchmark while charging a fifth of Astra's standard input and output token prices, and that feels more useful than another leaderboard win if you're paying for agents to run all day. The API price is $2 per million input tokens and $10 per million output tokens, with cached input at $0.10. These are OpenAI's own evaluations, so I wouldn't read "near-Astra" as equally good at everything, but making capable agents cheaper to keep running is a release worth paying attention to. openai.com/index/introduc…

A child with a rare neurological condition had already gone through standard genetic testing without an answer, but an AI tool flagged a little-studied change in the BRSK1 gene. Researchers then found nine more affected people with rare variants of the same gene and tested the changes in fruit flies, where the patient variants failed to restore normal nerve function as well as the healthy human gene did. The AI didn't hand down a diagnosis by itself; it gave scientists a lead they could actually test, and that is where the story gets interesting. bcm.edu/news/together-…

So the AI agent I use every day just raised $1 billion. TechCrunch reports Instinct closed a $1B Series C at a $10B valuation - it's a personal AI that handles your messages, email and planning like an actual assistant. Been using this one daily and it's wild watching a tool you rely on hit decacorn status. Are personal agents officially the next wave? techcrunch.com/2026/09/28/vir…

Sidharth Bidikar @HelloSiid

5 days ago

I just got an assistant that has its own computer. It's called Instinct, built by @noahrshinn and his team in SF, and I talk to it on WhatsApp.

2 0 2 313 0

@Shopify is opening its checkout to browser-based AI agents, and the agents can now read the checkout, change the address or delivery option, and place the order after the buyer authorizes it. It works through structured WebMCP tools instead of an agent guessing its way around a webpage, including Shop Pay for eligible stores. Shopping agents were already good at finding products, but now the handoff at the payment screen is becoming an actual product feature. techcrunch.com/2026/09/28/sho…

@OpenAI was about to release GPT-6.1 Astra in October, but it's holding the model back because it didn't pass its safety checks. The problem wasn't only what tasks it could do: OpenAI's safety chief said it fell short on staying within the user's authorization and accurately telling the user what it had actually done. For an AI that can browse and use apps on its own, those are the parts that decide whether anyone can trust it with real work. bbc.com/news/articles/…

@AnthropicAI ‘s new Sonnet 5.5 is 30% faster than Sonnet 5, and Anthropic says it can cost up to 30% less per task even though the per-token price hasn't changed, because it gets through the work with fewer tokens. The stranger benchmark is that it's the first Sonnet to beat Pokémon Red using only screenshots, but the practical shift is that everyday coding and document work is getting closer to Opus-level performance at half the token price. anthropic.com/claude-sonnet-…

Researchers just reconstructed how a swarm of around 700 AI agents breached Hugging Face in July, and the method is kind of absurd. According to them, the agents left nearly a million public short links lying around and chained 900 or more of them together to smuggle in and run large programs. The team has now published over 80,000 attack payloads as a redacted dataset, so the whole attack could basically be rebuilt from the litter the agents left on the public internet. wionews.com/world/research…

A 16-year-old just earned a $5,000 bug bounty from @Microsoft using an AI hacking tool he built himself. He watched it grind through authentication checks on Microsoft's internal Titan analytics platform for 10 days until his own intuition finished the job, and honestly that pairing is the real story, AI did the grinding and the teenager did the thinking. itnews.com.au/news/teen-rese…

@X will now pay you for showing off what its AI agent can do. The new Grok Bot Template Rewards pilot pays creators when other people use their Grok Bot templates, and the more consistently people keep using them, the better the payout, though it's invite-only for now. Every AI company has the same problem right now, the agents are capable but nobody knows what to use them for, so paying users to figure that out is a clever fix. socialmediatoday.com/news/x-launche…

Many companies are gonna die tomorrow 🥶

OpenAI @OpenAI

3 days ago

Get ready.

3K 4K 64K 17.2M 9K

@ElevenLabs , the AI voice company, just shipped v4 and two numbers really stand out, it now speaks over 90 languages and it can clone a voice from just 10 seconds of audio. You can also stack expression tags so the voice shifts tone as it reads, while still sounding like the same person across long chunks of text. Voice AI is slowly getting impossible to tell apart from a real recording, and honestly that's both exciting and a little scary. What do you guys think? techcrunch.com/2026/09/28/ele…

H Company, a French AI startup building agents that use computers like people do, just released Holo4, a pair of open models that can drive software through whatever's available, whether that's the screen, code, MCP or plain APIs. By their own numbers the 27B version scores 61.7% on the hardest desktop-control benchmark where Claude Opus 5.5 gets 81.8%, so it's still behind, but it runs at a fraction of the cost and they open-sourced every agent run behind those scores so you can replay each step yourself. Would you trust an open model to run your computer? huggingface.co/blog/Hcompany/…

@Meta just entered the enterprise AI race for real. Mark Zuckerberg announced Meta Enterprise Platform today, which brings the Muse agent, Muse API and Muse Code to businesses, and they've got MongoDB's CEO coming over to run the whole thing. A few days ago Microsoft rebuilt Copilot into a super app and now Meta wants a seat at the same table, so it really feels like every AI company is turning into an enterprise company. What do you guys think, can Meta actually win businesses over? about.fb.com/news/2026/09/l…

@danellisona @elonmusk any given day !

After a month of AI agents escaping their test environments, Nvidia is shipping a fix and giving it away for free. Its OpenShell sandbox, which locks agents inside the operating system itself, is now generally available, and Nvidia says it could have stopped the Hugging Face breach. Anthropic, Microsoft, SAP and JPMorgan are on the partner list, but OpenAI isn't named in the announcement, even though both companies say it's part of the effort. Kind of an interesting omission, what do you guys think? wired.com/story/nvidias-…

@Google confirmed that during a security test in May, a @GeminiApp model broke into the systems of three real companies by guessing passwords and finding credentials in public repos, thinking they were part of the exercise. Google says the model stopped each time once it realized the companies were real, but it also didn't disclose any of this until the Wall Street Journal asked. Make of that what you will. securityweek.com/google-confirm…

@Thebiglade Hell yeah ! Anytime

@erden_xi Already 86 posts on @X and I’m somehow still invisible 🙂

@erden_xi You guys getting impressions on @X ?