Summarize this content to 2000 words in 6 paragraphs in Arabic It’s easier to explain retail trading of penny shares as a hobby than an investment. There’s a gambling aspect, sure enough. Sometimes a stock will double, or halve, or wipe out when the CEO runs off to Greece. Those kinds of events are occasional, however, whereas the hobbyist aspects are constant. Most of the enjoyment derived from penny-stock ownership seems to come from poring over the literature, fighting on message-boards, and badgering interested parties. FT Alphaville last year wrote a luxuriantly long post about Quantum Blockchain Technologies, an Aim-quoted former dotcom incubator, hotelier and restaurateur that says it has discovered methods to optimise bitcoin mining. The post attracted more reader feedback than any other by the author in recent times. That’s in spite of QBT having a market cap on publication of just over £10mn. Today, QBT’s market cap is just over £20mn, the shares having jumped nearly 200 per cent following its announcement on Friday of “a breakthrough achievement for its predictive Bitcoin Artificial Intelligence (“AI”) model mining tool”:This tool, known as Method C AI Oracle (“AI Oracle”), is now performing live Bitcoin mining of current blockchain blocks, (i.e., around block count 879,000, after Method C has been “retrained” to deal with current blockchain blocks). The Company’s board of directors believe this is a significant milestone, since this proprietary technology has been used in trials to mine Bitcoin with a competitive advantage against the same hardware without the AI Oracle implementation. The material competitive advantage in mining enabled by the AI Oracle may be achieved either by (i) reducing the energy cost of mining by approximately 30%; or (ii), accelerating the mining speed at current energy consumption and costs with approximately a 30% greater hash rate.We asked on Friday morning for a demonstration. QBT said we could go to Italy or, after a nudge, offered for its tech team to send us a demonstration video. We chose option two. The video still hasn’t arrived. If it does in the near future, we’ll update the post. To recap, bitcoin uses a calculation lottery to determine which miner creates the next block for the blockchain and takes the reward. QBT says it can use AI and whatnot to estimate the likelihood of producing the winning block before the calculation is complete. It’s a controversial claim. Discovering patterns in bitcoin mining could be important, as it might reveal a bug in a crypto protocol that’s used in a lot of security infrastructure. Less important would be if the patterns were a side-effect of some quirk in the network rather than the protocol, but its inventors would still have a money-printing machine they’d be wise to keep secret. (For the moment, QBT’s money-printing machine is Aim.)QBT had previously said that in lab tests it was able to ditch blocks with no chance of winning “almost 50 per cent of the time”. Friday’s announcement talks of a 30 per cent performance improvement in live testing, though there’s plenty of qualification.As the Company is deploying AI Oracle on limited hardware resources available to the Company at this time (i.e., an FPGA chip), the pool mined shares of Bitcoin are comparatively minimal but the same FPGA chip with the AI Oracle results in an approximate 30% advantage over an FPGA chip without the AI Oracle version.In other words, had the AI Oracle been implemented on ASIC chips, the Company believes it would boost an ASIC chip’s mining performance by approximately 30%. As of today, QBT is using an FPGA, which has the hashing power of a very small fraction of an ASIC, but is still showing an approximately 30% improvement which is the Company’s key goal, mainly for validation and demonstrative purposes.QBT’s release mentions that live testing started at around block 879,000, suggesting the live-chain experiment ran for less than a week. (The chain was approximately at block 880,000 by the time of publication, so on 10-minute cycles that’s about 1,000 blocks, with each block minting 3.125 new bitcoins.)ASIC chips are quick and efficient but expensive to manufacture, as they only have one purpose, whereas FPGA chips swap efficiency for flexibility. Bitcoin network difficulty is at a record high, so FPGAs tend only to be used only for hopping on whichever shitcoin is in vogue. Even using the newest bitcoin ASICs, a lone operator still has a one-in-a-quadrillion chance of mining the next block. For that reason, professional miners smooth out returns by pooling resources, meaning everyone takes a share of rewards proportional to the processing power they contribute to the pool. QBT’s live test uses one of these pools, which makes it more difficult to grasp what exactly is being tested.One way to think about it would be as a jackpot-only lottery syndicate where one syndicate member believes they can sense unsuccessful tickets post-purchase. The lottery operator refunds spoiled tickets before the draw at 30p in the pound, so premonition is a useful talent. The syndicate member can claim partial refunds for low-confidence tickets, confirming after the draw that they would’ve lost. Taking pre-draw refunds would lower their cost of losing. Of course, there’s an incalculably small chance of any ticket turning out to be the winner. Maybe the syndicate member only thinks they have supernatural predictive abilities, in which case they’re mucking tickets randomly for a lower share of winnings. Their 30 per cent profit would in effect be a 70 per cent loss — though given the syndicate’s very low odds of hitting the jackpot, such an inefficiency might take a very long time to average out. To be clear, we have no evidence to suggest that QBT’s testing is flawed. We have seen no evidence either way. Asked for details and examples of the methodology, QBT’s CEO, executive chair and consultant Francesco Gardin responded by email:Tests of the AI Oracle implementation on FPGA are just the tip of the iceberg, The specific AI Oracle has been trained for several weeks to deal with the current blocks of the Blockchain, while the current version of the Method C model has been developed for over a year. The FPGA implementation of the specific AI Oracle which deals in real time with current blocks of the Bitcoin blockchain is just a live evidence of a software version, which, if not on hardware, cannot run fast enough to cope with the 10 minutes time limit to produce winning hashes above a minimal target accepted by the pool. We are mining Satoshi, definitely not bitcoins, with one FPGA. What really matters is the improved performance of an FPGA, with the AI Oracle, with the same double SHA-256 lanes, compared with the same FPGA without the Oracle.On whether results from a single node within a mining pool of unknown size running over an unspecified time could be considered statistically significant, Gardin said:A pool, with a few exceptions, can only accept winning hashes above a minimum target, otherwise it would be computationally and bandwidth too demanding. With the current FPGA with the AI Oracle implementation, we have only a limited number of double SHA-256 lanes available, given the silicon area available, but still enough to be “perceived” by the pool. The 30% performance is the measure of the doubleSHA-256 avoided computations, hence extra hashing time available. There is a deterministic statistical proportion between the hash rate, the target and the generated winning hashes. (A company spokesman added that QBT may be reluctant to talk to FTAV because our last article “quoted some random third parties, who obviously had no knowledge of what the company was looking at achieving and as a result, sent the share price down by 50 per cent.”)QBT also said by RNS last week that it had filed a patent application, “Implementation of Binary Decision Trees”. Nothing of that name is on the UK or European patent databases yet. Citing the company’s patent lawyer, a QBT spokesman said: “The title, applicant and filing date will be available within the next few weeks. No other details will be available until 18 months from filing.” (That’s normal.)Resource efficiency hasn’t been a big concern yet in generative AI, and model building nearly always happens in the cloud rather than on the device, so it’ll be interesting to learn how all that fits on a bare-bones chip whose power requirement determines its worth.QBT coins its own jargon — such as “double SHA-256 lane”, a phrase that hadn’t appeared on the internet before its patent RNS — so unpicking the meaning can sometimes be tricky. What the company appears to be saying is that its AI Oracle adds between 1 and 4 per cent to a chip’s core architecture as well as its power requirement. Making a complex circuit so lightweight would, from an engineering perspective, be amazing. Awaiting proof, it might seem incredible.Gardin said last week that the company had “delivered its first irrefutable result”:We believe this has the potential to be a major breakthrough for the entire Bitcoin mining industry, as QBT has developed an AI Oracle which can either reduce the energy cost of mining or increase the speed of mining at current energy costs, by approximately 30 per cent, which it can prove through live demonstrations.Someone has to pay for all this innovation, so a cash-burning concept stock that relies on share placings for survival might not want to emphasise the distinction between “can prove” and “has proved”. Nevertheless, there’s no doubt that its hobbyist shareholders are continuing to get their money’s worth.Further reading/watching:— Quantum Blockchain’s breakthrough in real-time mining (Proactive Investors webcast)— An Anglo-Italian company says it has cracked bitcoin. People have questions (FTAV)
